LiteSpeed Hosting as
Cloudways Alternative

Part 3

  • Select the hosting plan with the LiteSpeed Web server and NVMe SSD storage
  • Choose a shared server geographically close to your visitors
  • Don’t go for the Cloudways hosting services – it is the worst value for money
  • Study the list of standard software suite which is provided in this report
  • Recognise that in the current market, the hosting provider size is its liability
  • Further analysis of the hosting providers can be found in our separate report
  • In the US or Europe, host your website with MechanicWeb – the best value provider
The article takes 28 minutes to read
Last modified on November 6th, 2021

This post is a part of our main research article

We wrote our main research article to inform a large community of web designers and business owners about the best inexpensive options for shared WordPress hosting. Our goal is to help you optimise asset delivery to achieve the best Core Web Vitals score by combining fast hardware, advanced security, and optimal caching.
Different WordPress groups on Facebook have the perfect symbiosis and coordination. This is teamwork. Some “experts” will try to convince you to switch to the “best” shared hosting plan with slow hardware, while others will try to convince you that there is nothing faster and more secure (sic!) than an unmanaged Vultr HF VPS on Cloudways. In both cases, affiliate fees are collected by misleading the unsuspecting public and convincing them to switch to the wrong service provider.
Many users are surprised to learn that there is a third option: Premium Shared Hosting providers with hardware as good or better than unmanaged VPSs. With cPanel, server-level malware protection, and LiteSpeed Enterprise, you get a much more secure experience and faster website loading speed from Premium Shared Hosting. For WooCommerce sites, the benefit is even greater. The best providers may not have an army of affiliate warriors behind them, but that does not mean their services are any less attractive.
Our main research article provides a quick overview of the best hosting options currently on the market. The following posts continue our research article.
  • Part 1. We give a detailed overview of the standard software options offered by all reasonable LightSpeed web hosting providers. Before you choose your hosting, you should read this chapter!
  • Part 2. You might read the detailed chapter explaining the cheating and fraudulent schemes of Cloudways and its affiliate community.
  • Part 3 (this post). We explain the technical details of our benchmarking study. Read it if you want to quickly test the actual quality of hardware resources provided by your hosting provider.
  • Part 4. Covers advanced hosting options for extremely busy websites. You should study this chapter if you are determined to switch to VPS hosting.

Our Methodology

The free plugin WPBenchmark provides you with comprehensive information on server performance. The plugin is listed in the official WordPress Plugin Directory. Please use default WordPress plugin installation method through WordPress plugin repository.
WPBenchmark
With this plugin, you get everything you need to measure by simply activating and running it. In about five minutes you will see results similar to the one shown below.
Results for MechanicWeb, Ryzen 9 5950x Zen 3 server
MechanicWeb_WPBenchmark
A2Hosting_WPBenchmark
Results for A2Hosting, AMD EPYC 7302 EPYC server
So you can understand the results:
  • “Operation with large text data” is your CPU score; you should aim for less than 6.5 seconds. Best value is 3.5 seconds.
  • “Random binary operations” – this is your CPU’s math performance; if it is over 6 seconds, you should run away. Best value is 3 seconds.
  • “Filesystem” section – if you see any of these numbers below 30 MB/s, run away; otherwise, ignore it.
  • “Importing large amount of data to database” – indicates whether your hosting provider needs to hire another sysadmin; run away if you see numbers over 5 seconds. Best value is 1.1 seconds.
  • “Simple queries on a single table” – again, indicates whether your hosting provider should advertise a new position for a sysadmin ; values over 1.6 seconds are not acceptable! Best value is 0.4 seconds.
  • “Complex database queries on multiple tables” – should take less than 0.7 seconds. Best value is 0.2 seconds.
  • “Network download speed” – avoid the 24 mbit/s that many old-school providers offer; you deserve better!
With that simple methodology, you can conclude that MechanicWeb’s offer is the best, and that A2Hosting does not deserve your attention.
A mild warning. Please do not use the WPBenchmark plugin frequently on production servers. It is a resource-intensive test. If enough users start to perform those benchmarks repeatedly, it would hike up the server load and might force a provider to prevent that situation. Be a responsible user – a one-off test is more than enough!
You can further test your shotlisted provider by measuring TTFB for your non-cached site after installing vanilla WordPress. The test location near the server should have no more than 300 milliseconds of latency. KeyCDN’s Performance Test tool is the easiest way to measure TTFB. Below you can see the results for MechanicWeb’s servers in Germany. It shows non-cached performance for the dummy website using a fresh installation of WordPress, Elementor, and the Hello Elementor Theme.
MechanicWeb_Hetzner_uncached
Repeat the above test with the cached website. You should see the value fall below 60ms. If your TTFB for the cached website at the location near your server is over 180 ms, run away. This concludes the test. Congratulations to your new provider if they pass your tests!
In the remainder of this Chapter, you will find details of the research validating the procedure described above. We ran three alternative benchmark tests, as well as the K6 loader, to test site performance with 100% loaded CPU on non-cached sites. We discovered some interesting anomalies with the K6 loader. It turns out that many companies limit in undocumented ways the maximum number of processes a user account can run. WPX has two Nproc configured, HostMantis has five Nproc, Closte uses ten Nproc, and WPEngine uses thirty Nproc.
Raidboxes’ performance was the most shocking result. The company is known for rewriting WordPress core files to improve cache performance. Tests have shown that their un-cached performance is degraded by more than three times. They use old, slow hardware, so their performance without cache is about 9 times slower than the top performer in our test, MechanicWeb, which uses Ryzen Zen 3 servers.
We created a dummy website for TTFB and load testing using a fresh installation of WordPress, Elementor, and the Hello Elementor theme.
Below is a graph comparing the results of the K6 load test and the performance rating of CPU. Most hosting providers show a shockingly good correlation. NameHero and A2Hosting use the same AMD EPYC 7302 server hardware, so their data overlaps. They overlap even though they have very different database access times and disk I/O speeds. The result for Hostinger is not negatively affected by the disproportionately slow disk I/O speed. This is to be expected since a K6 load test was performed on a vanilla WordPress installation with Elementor and Hello Elementor Theme. The website is too simple and the database is too small for these parameters to be a bottleneck to actual performance.
K6_loader_vs_CPU_score
 Data from the K6 loader was normalized by the number of CPU cores in the plan. The main “blue” group shows a surprising correlation with the CPU performance score.
In the “orange” group of hosting companies, you’ll find EthernetServers, A2Hosting, NameHero, and GreenGeeks. Google N2 instance from another hosting company, Stablepoint, completely overlap with A2Hosting and is not shown in the Figure. Another server from Stablepoint, rented from Vultr in Singapore,  almost completely overlap with EthernetServers. All benchmarks of the hosting hardware of those providers has one thing in common. According to the WP Benchmarking Plugin report, their servers have a pathetically slow rating for “Operations with data”. Their hardware is extremely slow not only for math operations but also for almost all PHP functions.
A separate study of PHP benchmarks confirmed that the EPYC 7302 servers used by NameHero and A2Hosting perform only at 66% level of Brixly’s EPYC 7502 servers. The K6 load data from Brixly and NameHero correlated very well with this result. The Intel Xeon 2.8 GHz from Google Cloud used by Stablepoint scored 2% higher in the PHP benchmark test than AMD EPYC 7302 used by NameHero. Satblepoint benchmarks overlaps with AMD EPYC 7302 in most other tests shown in our benchmark tables. EthernetServers’ PHP performance metrics are on par with Guru’s, but its actual K6 load test scores are about 25% lower. This can be attributed to a much lower score in the string_simple_functions metric.
We should emphasise that AMD EPYC 7302 (used by NameHero and A2Hosting), E-2136 (used by EthernetServers), and Google Cloud N2 (used by Stablepoint) are not suitable for WordPress websites that require a lot of processing power. Such servers should be preferably used for static, fully cached brochure websites.
Data for Raidboxes are the only ones in the red group. Raidboxes rewrite core WordPress files, and their results show that software bugs can completely sabotage hardware performance.
Times are changing. Any hosting company using Nginx needs to configure it properly to perform well for cached content. As a result, there was a need to measure the actual performance of your server when serving cached content. With LiteSpeed, the situation is different. LiteSpeed is better than Varnish but much easier to manage and tune. LiteSpeed provides perfect cache performance even on old, slow hardware. It has the same performance as a local CDN node, making LiteSpeed the ideal choice for many European businesses that are leery of CDNs due to the GDPR. The graph below shows LiteSpeed’s performance with a cached website on a Ryzen Zen 3 server. It can handle 600 visitors per second, or two million per hour, with a TTFB of 65 ms. Still, this load is far from LiteSpeed’s limit!
MechanicWeb_LSCache
Loarder.io – fully cached website on MechanicWEb servers.
We have a simple message:
  • Choose a hosting provider with modern hardware, like AMD EPYC 75xx, Intel Xeon E-2288G, E-2388G, or AMD Ryzen 9 5950X.
  • Make sure you are served with cPanel and LiteSpeed Enterprise.
  • Please make sure you get a firewall and malware protection set up at the server level.
  • Run the WPBenchmarking plugin to determine the credibility of the sysadmin team; pay attention to the Filesystem and Database sections.
  • Enable the Quic.cloud CDN.
  • Use AssetCleanUp and learn how to configure the LiteSpeed Cache plugin.
  • Learn about the Kadence Theme and Kadence Blocks, Bricks Page Builder, and how ESI integrates with WooCommerce.
  • Get the highest Core Web Vitals rating possible.
The above information is all you need to know about optimizing your website speed and choosing a suitable hosting provider. You can jump directly to the discussion of results of individual provider or just enjoy a summary Figure below.
LiteSpeed_vs_Cloud
For more details on the LiteSpeed web hosting providers examined in this price-performance analysis, see our separate report.
Our approach to testing hosting providers is based entirely on the useful WPBenchmark plugin developed by Anton Aleksandrov. We have asked Anton to change some aspects of his plugin to ensure it can run in the most adverse hosting environments, where some of his tests will fail. This is the only value we bring to his plugin. We believe that our research analysis is a more important contribution to the validation of this plugin as a benchmarking tool. You should only read our study if you can handle technical details, otherwise it is too hard to digest.
Our research was met with a wall of aggression, hostility, and personal insults from many senior members of the WordPress Hosting Facebook group. Hosting “experts” cannot bear the notion that any web designer can benchmark the hosting providers with a few clicks of a mouse. Over the years, the hosting “expert” community has developed mythologies about the complexity of testing hosting providers and the need to become a community certified expert to understand any aspects of proper performance analysis. It is an elite sport!
These religious beliefs attract affiliate scammers and plain charlatans, who commonly claim to test uncached website performance but leave several cache layers, such as OPcache, Nginx, Varnish, or even CDN untouched for “privileged” hosting providers. Their motto for priveleged clients is “Don’t worry, no CPU will ever be stressed by our testing!”.
It is highly regrettable that most hosting “experts” fail to understand tricks and to expose fraudsters and cheaters. They find it even more difficult to accept that outsiders now have access to the free tools and can outwit them when testing web hosting provider performance.
We believe our research plays a vital role for the community by clearing the mental fog caused by the generation of affiliate scammers and fraudsters. We provide the details of our study, which was conducted using only free tools, and invite anyone who would like to review our numbers for specific providers. The hosting plans listed in our spreadsheets can be opened and closed with a full refund. To check our numbers, you do not have to invest anything but your time. In the following sections, we present our data with the belief that our research provides comprehensive answers to most questions.

WPPerformanceTester plugin

WP Benchmarks

The original publication from Matthew Woodward states, “All previous tests were served from cached results, but the WordPress Performance Tester plugin was designed to test the website performance of the actual server itself for each WordPress site.”
WordPress based benchmark tools is exactly the missing link in testing the server! Currently, we have limited number of free tools to measure the performance of the server. WP Performance Tester plugin, however, is one that general users should avoid! We used it extensively during our research study and highly recommend it to web hosting providers who want to quickly test the latency improvement during fine-tuning. Kevin Ohashi developed this plugin by combining open-source scripts that were publicly available. Unfortunately, he forgot to replace one of the functions in the script and now his plugin gives huge 75% advantage to AMD CPUs.
Hoping to spread objective benchmarks instead of unsupported marketing claims, he also put his WP Performance Tester plugin in the public domain. In addition, Kevin Ohashi invites WordPress hosting companies to pay him for the privilege of participating in his annual testing. Taking part in such annual surveys, using the WP Performance Tester plugin among other tools, is a regular occurrence for Cloudways. Therefore, Cloudways supports the use of this plugin. It is amusing to see that AMD Ryzen Zen 3 shows benchmark performance 3.9 times (!) faster than Vultr HF with WP Performance Tester plugin, even though Cloudways boasts of providing the fastest service to its customers.

Problems with WP Performance Benchmark

We installed the WP Performance Tester plugin on dozens of dummy websites to measure PHP benchmarks on different servers. The results of our study were compared with those of Kevin Ohashi from 2021 and data from Matthew Woodward. The plugin performs two types of tests. The “Server Performance Benchmark” tests the hardware capabilities of the server to run PHP code. The second, called “WP Performance Benchmark”, tests the WordPress interface’s latency when interacting with a database via the $wpdb class, bypassing SQL.
Based on our data and data presented by Kevin Ohashi, we found no correlation between hardware performance tests and latency tests. Hardware benchmarks correlate well with changes in CPU speed between different servers. However, the latency of interacting with a database differs significantly between different servers. It is not affected by caching, CPU speed, CPU cores, amount of RAM, or disk I/O speed.
WordPress Performance bencmark measures the latency of requests to the database. It is not related to the hardware performance of the server.
WordPress Performance benchmark measures the latency of requests to the database.
It is not related to the hardware performance of the server.
The data from Kevin Ohashi show no correlation between the two components of the test, while the data from Matthew Woodward show a near-perfect correlation. Consequently, we have found that Matthew has completely falsified his data. It’s not the first time people have published fake data, but it’s revealing to discover the obvious fraud in the data published by the well-known SEO blogger.
We expressed concerns about the WP Performance Benchmark part of the plugin tests in our conversation with Kevin Ohashi. We advised him to change the test title. It is puzzling that the WP Performance Benchmarks has changed twofold on Cloudways plans on two separate instances with DO Premium 2GB.  However, the Server Performance Benchmarks were identical. The problem is not isolated. The WP Performance Benchmark varied by a factor of 2.5 across three types of servers provided by MechanicWeb from Los Angeles, Dallas, and Germany, implying a latency variation of 0.6 ms to 2 ms.
However, MechanicWeb discovered the cause of this latency variation and reduced it to almost zero. In one of their data centres, WP Performance Benchmark had changed from 440 to 2801, more than a six-fold increase! And now it’s two times better than the best score from Cloudways, giving MechanicWeb the top spot among all other hosting providers participating in Kevin Ohashi’s annual surveys. Of particular note are the excellent results that MechanicWeb plans achieved in the Server Performance benchmarks. Their results were four times better than those of Vultr HF on Cloudways and two times better than the best results ever recorded by Kevin Ohashi.
Were there any improvements recorded by MechanicWeb after fine tuning? The largest change was for TTFB, which was reduced to 68ms. The overall loading time for a dummy website dropped by 150ms. The fine-tuning unleashed the hardware performance by removing delays in processing requests to the WordPress database. MechanicWeb was pleased with the test plugin as it allowed them to identify the problem and gave them a measurement tool to check for the bottlenecks on the server to fix the problem. The optimised server is capable of serving larger website traffic. This was confirmed in load tests using K6.io and Loader.io testing tools. However, WP Performance benchmark results cannot quantify the extent of improvement. It will very much depend on your specific website. We believe that a more holistic test such as the K6 load test for the uncached dummy website described below gives a better measure of the server’s overall performance.
The real-world performance of WordPress is not as severely affected by latency as implied by benchmarks. On servers with heavily throttled disk I/O speeds like A2Hosting with its pathetic 4 MB/s, only a meagre 2.5 kB of data is read or written in the 0.6 ms time span that corresponds to its database request latency. We don’t want to put that much emphasis on the exact values of the short latency of the database requests because it’s only one part of a more complex picture. This score surely provides a great indicator of the sysadmin team experience in fine-tuning WordPress performance settings, but it does not go much further.
The WP Performance Tester plugin provides usable results in the WP Performance Benchmark section to help a hosting provider fine-tune their server. However, it is not as useful to us for comparing different hosting providers. By measuring the latency of interaction with the WP database via the $wpdb class, the WP Performance Benchmark does not provide insight into the average performance of hosting providers in daily use. Using such a test to evaluate the WordPress software stack is not appropriate. Delays can be caused by settings in CageFS, imunify360, ModSecurity, MySQL, etc. Depending on the design of the benchmark test, delays may affect benchmark results differently from how they affect your site’s performance on dynamic queries.
Based on our analysis of the raw data, we have concluded that the second part of the plugin test, the WP Performance Benchmark, is an important but indirect and uncertain contributor in comparing hosting providers. Consequently, we use it mainly to benchmarks the quality of the sysadmin team in fine-tuning MySQL and other aspects of the hosting.  Unfortunately, the first part of this benchmark plugin,  the Server Performance Benchmark, has its own bug.

Problems with WP Performance MySQL test

Upon closer inspection, we found that the so-called MySQL query benchmark test is using a specific internal MySQL function and is heavily biased in favour of the AMD architecture. Since not all real-world processes are affected in the same way, this benchmark has its own limitations when comparing Intel servers to AMD servers. To establish the actual degree of the bias, we have carried out a cross-comparison of results measured using WPPerformanceTester plugin,  WPBenchmark plugin, and original Benchmark.php script. Kevin Ohashi adapted this script for his WPPerformanceTester plugin. Authors of the script have changed specific MySQL function used by the test in 2019. Kevin has ignored this change.  The results of the WPPerformanceTester plugin’s MySQL query test are shown below in comparison to the aggregate CPU benchmark, which is the result of averaging three independent benchmarks.
Benchmark_Intel_vs_AMD
While all the Intel-based vendors have fallen into a neat line, the four AMD-based hosting providers have taken a different path. With our knowledge of the exact server specifications used by the vendors, we can surmise that their aggregated CPU benchmark scores are correctly position them together with the Intel-based servers with a similar CPU frequency. This chart illustrates the significant, 75% performance gain for AMD when it comes to specific MySQL function selected by Kevin Ohashi.  In real-world applications, however, these benchmark results should not be trusted without authoritative investigations. For our combined benchmark tables, we multiply the WP Performance Tester MySQL test results for AMD-based servers by 1.75 to correct for the artefact.
Please be aware that AMD EPYC performance boost is not universal for all MySQL functions. The WPPerformanceTester plugin uses the ENCODE() function to measure the so-called MySQL query test, which gives an estimated boost of about 1.75x to AMD. Using the AES _ENCRYPT() function of the original MySQL Benchmark.php test removed the boost for AMD servers. The figure below does not show a breakdown by Intel and AMD architecture. This means that the boost for AMD servers is an artefact built into its hosting benchmark plugin by Kevin Ohashi. In other words, please be aware that benchmark plugin developers don’t care about performing thorough testing and analysis of their algorithms.
Kevin Ohashi deserves some criticism, followed by praise. The WPPerformanceTester plugin is an indispensable tool for testing hosting providers. With it, you can quickly tell if you are dealing with AMD or Intel servers. Without that significant artefact generated by Kevin Ohashi, it would be almost impossible!

WPBenchmark and Benchmark.php tests

We have modified original  Benchmark.php  file – an open-source script which was copied and changed for WP Perfomance Tester by Kevin Ohashi. We have replaced the original string with the:
$query = “SELECT BENCHMARK(10000000, AES_ENCRYPT(‘hello’, RAND()));”;
We increased the number of tests tenfold and added the RAND() function. Our goal was to get Benchmark.php to perform a measurement as similar as possible to that of the WPPerformance Tester plugin. The only difference between the two tests now is the functions AES _ENCRYPT() and ENCODE().
Benchmark_PHP
We used the result of the “operations with large text data” test in the WPBenchmark plugin as a third cahnnel for aggregated CPU performance. The Benchmark.php tests appear to accurately measure the performance of CPU, but the WPBenchmark test results are slightly skewed by the software delays inherent in a given server.
WP_Benchmark
We corrected the AMD test results measured with the WPPerformanceTest plugin and calculated the aggregated CPU benchmark value as an equaly weighted average of the three benchmark channels. The aggregated CPU benchmark was normalised to the value measured for the Vultr HF server ordered from Cloudways. The hosting provider CPU scores are expressed as deviations from a Vultr HF reference value in percent. We calculated that the CPU score of the Ryzen Zen 3 server is 92% higher than that of the Vultr HF server. Raidboxes servers have a 28% lower CPU score than Vultr HF. Compared to Raidboxes servers, Ryzen Zen 3 server is (1+0.92)/(1-0.28)=2.67 times faster.

Independent AMD EPYC tests

Our conclusion about unqalified boost for AMD servers was put through its paces with a Google search to see if it stood up to professional comparison. We were pleased to find a recent report, “CPU performance in shared hosting environment”, from the CloudLinux team that examined the same issue. To avoid similar embarrassments shown with our WP Performance Tester plugin analysis, the CloudLinux team created a professional representative mockup of the shared hosting environment with CageFS. They put 200 user accounts on each server, installed WordPress for each user, and created 12 unique pages on each domain.
As a result, the research team obtained 2,400 unique URLs to test server load. Random requests were sent to 2,400 URLs. It took 5 minutes to complete a test iteration. Each iteration had 150 concurrent, non-keep-alive connections. That is approximately two concurrent users per logical CPU core. At the end of each iteration, the total number of requests processed and the number of requests per second were calculated. The spread between different servers was between 12 and 16 requests per second per logical CPU core.
CloudLinux team concluded that the AMD Rome (Zen 2) architecture offers from 0% (for 7302) to 18% (for 7542) better aggregated CPU performance, while AMD Milan (Zen 3) offers 28% (for 7313) more CPU performance.  These numbers are derived by correcting the raw data by the actual CPU speed of 2.9 GHz for Intel Xeon and 7542 vs 3.0 GHz for 7302 and 7313. The researchers also measured a 9% performance boost for LiteSpeed web servers in their benchmarks for web servers and PHP handlers. The numbers above are significantly lower than those claimed by the LiteSpeed team, but this integrated metric provides a more accurate comparison. The researchers found that the new AMD EPYC offers a 32% lower price per request, suggesting that more hosting providers switch from Intel to AMD EPYC in the coming years.
AMD gain of function
Maximum requests per seconds on 32 CPU core servers
with the same frequency on CPU and RAM
as measured by CloudLinux team

K6 load test - a part of our benchmarks

We created a dummy website for TTFB and load testing using a fresh installation of WordPress, Elementor, and the Hello Elementor theme.
We briefly used Loader.io to test performance of up to 2,000 visitors per second on a regular shared hosting account running LiteSpeed Enterprise on Ryzen Zen 3. TTFB stays below 60 ms. This number is completely meaningless to 99.99% of users. You should be interested in a stress test that uses your entire CPU quota. So you need to test non-cached websites. A 5 minute stress test with the K6 loader is better for this purpose.
Running free tests provided by K6 is the easiest way to reproduce the CloudLinux trial described above. A dummy website that can be replicated by anyone is created by installing WordPress, Elementor, and Hello Theme. The web pages should not be cached. This forces the server to compile PHP code and provide HTML, CSS and JS code after extracting the data from the database. Anyone can rebuild such a dummy website and run up to 50 tests with a free K6 account. We used 5-minute tests where the number of virtual visitors was ramped up during the first minute and ramped down during the last minute of the test.
K6_get
K6 settings
Setting up your test in the K6 loader is simple. Under the Requests – GET you specify https://testdomain.com. That’s all you need to do, aside from changing the number of virtual users (VUs) under the Ramping VUs section.
Because Cloudways does not allow OPcache to be disabled, their scores for this test were higher than they should have been for a fully comparable test. OPcache typically increases the results of such tests by a factor from five to ten. To reduce OPcache’s efficiency, we reduced its memory limit to the minimum allowed value of 9 MB. Nevertheless, it is reasonable to assume that the Cloudways results are overestimated.
By changing the number of virtual visitors, the average response time per visitor can be increased to over one second. Under such conditions, the number of virtual visitors served per second reaches a saturation point. As the number of concurrent visitors increases, the response time per visitor increases, and as the number of simultaneous visitors decreases, the response time per visitor decreases. However, the system throughput (number of visitors per second) remains the same. Such saturated system performance – the number of visitors per second – is proportional to the CPU allocation. The cPanel resource monitor shows that under saturated conditions, the K6 tests use 100% of CPU capacity.
K6_test
K6 load test for MechanicWeb, 2 CPUs plan. The sky blue line indicates response time. It should increase to over 1 second once the number of virtual users reaches the maximum value in the test. Modify the number of VUs to bring the responce time into the range from 1 to 1.5 seconds.
In the following graph, we compare the performance with the monthly fee paid to the hosting providers. The price-performance ratio can be easily seen from this graph. There is no doubt that LiteSpeed hosting providers are significantly better than Cloudways. Ten times better value for money is no small thing compared to the claims of Cloudways affiliate scammers!
LiteSpeed_vs_Cloud
For more details on the LiteSpeed web hosting providers examined in this price-performance analysis, see our separate report.
Some hosting providers have failed the K6 loader stress test. Examples of this are Closte and WPX. WPX is particularly affected. They have oncet of errors and spikes in the response time even when serving only two visitors per second. Please refer to the Figure below for details. At a higher visitor rate, the errors become dominating. WPX is limiting its $21 monthly plan to just three PHP workers. Our data implies that you should count for only two PHP workers. It looks so restrictive and completely overpriced! Have you ever seen a report from Kevin Ohashi about such problems at WPX? No, because he’s not interested in doing a proper stress test for non-cached sites. At the very least, you need to spend some time removing caching layers and checking that you have done that. Caching remains with many providers. KnownHost, Nexess, Krystal Hosting, and WPEngine are just a few examples.
WPX
The results measured for Closte are very similar to those measured for WPX. The Closte plans are limited to 10 Nproc, so the errors occur at much higher loads.
Closte_K6
HostMantis demonstrated something very intriguing. Because of an undocumented limitation of only 5 Nproc per account, K6 load tests with the 2.5 CPU plan only used up to 1.6 CPUs. The screenshot below also shows the onset of errors and spike to 4 seconds for the response time. This is almost identical to what we saw at WPX and Closte, but now we also have access to the CPU load data. HostMantis appears to be limiting their $17 per month plan to just five Nproc while showing the limit of 300 Nproc under cPanel. Our stress test displays a lack of professional sysadmin skills at HostMantis. Unfortunately, we were not able to discuss these results with HostMantis, which just refunded our account citing that performance testing was against their Terms of Business.
HostMantis
K6 HostMantis
In the K6 load test, the Raidboxes showed pathetic results. During a stress test, their plan spiked to 800% CPU utilisation on a plan limited to one CPU – please refer to the Figure below. All 8 cores were serving just 6 visitors per second on average, and thus we can state that Raidboxes is able to serve 0.75 visitors per second per CPU. Ryzen Zen 3 can serve more than seven visitors per second per CPU. Or, to be more precise – ten times more visitors per CPU than Raidboxes. It is absolutely safe to approximate that all dynamic uncached requests will be processed on Ryzen Zen 3 server ten times faster. Does Kevin Ohashi have anything to say about Raidboxes? He put them in the premier league! Do not think Kevin Ohashi’s results are the gold standard.
Raidboxes_CPU
Raidboxes_K6
So why is Kevin Ohashi recommending a junk hosting provider like Raidboxes as the best option to his unsuspecting readers? It has to do with money. See below how much Raidboxes pays for scammers who recommend its services:
They can earn between €75 and €1500 for each referral. The commission is five times the monthly fee of the referred customer.
Cloudways pays up to $125 per referral, regardless of user plan. And you’ll have to compete with an army of affiliate scammers looking to promote the same compromised Cloudways services. With Raidboxes, your chance of earning higher affiliate fees is even greater! Irresistible, is not it?
A 1vCPU plan on Raidboxes spikes up to 8vCPU occasionally, so many testing results are too positive to be true. Their actual hardware is nothing too extraordinary.
Their caching is great, and I have been fairly impressed by load tests, but be assured that support will contact you if you’re actually overusing your plan. But I have also heard about a lot of issues with WooCommerce, or in general dynamic sites, and their caching.
In K6 load tests, Ryzen Zen 3 served two times more visitors per CPU core per second than Vultr HF. These tests were repeated multiple times with no errors.

Results of our Benchmark study

We created a dummy website for TTFB and load testing using a fresh installation of WordPress, Elementor, and the Hello Elementor theme.
Faster website performance can be worth more money in some cases. What then can be found out about the speed difference between different hosting providers? As discussed in the previous chapter, we use two free WP plugins, such as WP Benchmark and WP Performance Tester, to run tests. Additionally, we used Benchmark.php – an open-source script which was copied and modified for WP Performance Tester by Kevin Ohashi. Benchmark.php was used because it uses MySQL functions that do not favor AMD processors. The KeyCDN Performance Test was also used to test TTFB at the geographic location closest to the hosting server.
Top_LiteSpeed_providers
The table lists some of the best LiteSpeed hosting providers. The “MySQL function” test for AMD servers has been corrected by multiplying by 1.75 to remove artificial bias; these data are marked with the asterisk.
Top_Nginx_providers
We do not recommend these hosting providers.  We just want to show you benchmarks of the best Nginx hosting services. Cloudways’ test results are overstated because it is impossible to turn off OPcache; we reduced the memory allocation to the minimum allowable value of 9 MB for OPcache to reduce the boost. Raidboxes is inefficient and slow at delivering non-cached requests; we question the usefulness of using such a service.
All-incl, Nexcess (Liquid Web), SiteGround, and 20i do not limit CPU allocation, so the K6 test does not show saturation for the number of visitors served. SiteGround limits your account by total monthly visitors, while 20i limits it by the overall monthly bandwidth. WP Engine shows highly dissapointing results. Their caching layer cannot be removed, but WP Engine will start rate limiting your site if the number of visitors exceeds 30 per second.
Second_Tier_LiteSpeed_providers
To help you choose the best LiteSpeed hosting provider, we have benchmarked more companies. We do not recommend any of the hosting providers listed in this third table. At maximum load, Closte generates errors. WPX generates errors and experiences sporadic delays of TTFB up to eight seconds at just two visitors per second. Benchmarking studies on Hostinger, and GreenGeeks show them to be pure junk-grade providers, yet they are enthusiastically recommended by a large number of illiterate affiliate scammers.

The importance of bypassing caching

Most LiteSpeed vendors provide access to PHP extensions through the cPanel, so caching can be easily configured. However, with many of the companies compared above, caching settings are hidden behind proprietary control panels. We always added the code recommended by A2Hosting to the .htaccess file to turn off hidden caching levels on the Apache webserver. If you don’t have access to the files on the server, you can use LiteSpeed pluging to edit the .htaccess file. Alternatively you can use WP File Manager plugin.
In some cases, such as WPX, this works well. Some caching levels, like OPcache, remained active in other cases, such as KnownHost. Based on the K6 load test result, we can see how the response time increases with the number of virtual visitors served by the K6 loader. If this time does not increase under heavy load, it can be interpreted as a sign that some caching layers are still active. The number of users served per second should be saturated at a reasonable level while the CPU runs 100%. This is a situation where you can be sure that you measure the uncached capabilities of the server.
It is essential to recognise how our tests differ from those of uneducated affiliate scammers. We take extensive steps to disable caching layers and confirm their removal. We test server performance instead of external caching layers like Varnish servers or CDNs. K6 load and TTFB tests can give you an idea of server performance on dynamic requests for non-cached resources. LiteSpeed server performance when serving fully cached resources is comparable to Varnish cache. On LiteSpeed, the TTFB values for cached resources are even lower. Whenever we fail to disable caching on a particular server, we record it. We are not telling you that this server is better than any other on the market. Not at all! In its determination to prevent us from changing caching settings, this hosting provider outperforms all of its competitors. That is the only achievement of this provider!
On LiteSpeed servers, brochure websites are served directly from the server cache. They are delivered quickly, even if your server is slow. Actually, they are delivered as fast as from the CDN node. Fast hardware speeds up the Elementor editor and other tools you use to edit your site, but it’s not critical to the static website performance. Any website with dynamic content, such as WooCommerce, Magento, or BuddyBoss, should be concerned about the speed of delivery of their non-cacheable content. Since the vanilla website used in our test did not create enough issues with slow database accesses or slow disk I/O speed, our K6 benchmarks for non-cacheable content can not provide a complete picture. However, it clearly shows the importance of fast performance from the CPU.

Our findings

Unsurprisingly, Cloudways is not the top performer in our benchmark tables. It’s just as good as Brixly, GURU, Kinsta, Nexcess, and SiteGround. The performance of companies like Nexus Bytes, EthernetServers, NameHero, Nexcess, WP Engine, 20i, and Raidboxes is degraded due to unoptimised MySQL settings. As a result, these companies appear to offer the worst performance scores in the “Database” and “WordPress Performance” sections. The culprit here is an unprofessional sysadmin team so you should think about what else might be wrong with the security settings of the compromised accounts. Some of the companies above don’t have in-house sysadmins and rent resources from a Data Center, who configures high availability cloud shared servers, the CloudLinux, and CentOS 8 installation themselves.
Sewer.io, Brixly, and MechanicWeb are the winners in the speed test. Using modern servers helped these companies to the top place. MechanicWeb and Brixly use Zen 3 generation servers, which are more than 45% faster than Serwer.io’s Zen+ generation servers. MechanicWeb’s Intel Xeon servers are only 10% faster than Serwer.io Zen+ servers. Server performance has been radically improved with AMD Ryzen Zen 3. One of the first hosting providers to benefit from this significant increase in server speed is MechanicWeb. Brixly has joined them by renting Ryzen 9 5950x servers in the same Hetzner data centre in Germany.
Kevin Ohashi’s credibility was further undermined by a terrible performance by Raidboxes plan for providing uncached resources. In his latest review, Kevin Ohashi recommended 20i, GreenGeeks, WPX, and Raidboxes as top-notch shared hosting providers. They all pay affiliate fees of about $100 for each new paid customer; Raidboxes pays even higher fees. If you take affiliate fees out of the equation, this list sounds like the bottom of our benchmarks. We would definitely put Raidboxes at the bottom of any list, no matter how junky it is.
Kevin Ohashi claims that his LoadStorm test measures uncached performance. In his own words, “Uncached load testing is requests that bypass caching and hit PHP and the database. Logging in a user satisfies this by definition, since it has to authenticate with the database, bypassing all static caching layers.”
It all sounds very nice, but:
  1. LiteSpeed can cache pages for logged in users. The new Guest mode is designed to deliver cached pages even on the first visit. In Kevin’s current method, this is not accounted for. Be careful – many LiteSpeed hosting providers recommended by Kevin have very old slow hardware.
  2. Kevin is aware that his WPX test is skewed by measuring their CDN performance. CDN is not a web hosting product. When comparing web-hosting providers, it still matters whether you measure CDN performance or the server performance of WPX. Our data shows that WPX will fail a fair test.
  3. The LoadStorm test puts providers that do not enable OPcache by default at a significant disadvantage. OPcache reduces server load by about 10x for a simple un-cached website. Benchmarking reviewers must discuss this issue and allow or disable OPcache on all accounts tested.
  4. No data is collected on CPU usage or the number of processes active during the test. This would undoubtedly shed light on the validity of the LoadStorm test and expose the hardware speed of the servers. Not only for the audience but also Kevin Ohashi himself, the test is a black box. In the same way that Kevin does not comprehend what is measured by his plugin, he does not understand what is measured by his LoadStorm test.
  5. Raidboxes explicitly allow their 1 CPU plans to spike up to 8 CPU during the test. This is a typical Dieselgate method, designed only to pass tests like Kevin Ohashi’s LoadStorm. Kevin Ohashi claims that it is important to him to detect any form of cheating. But why is Raidboxes cheating allowed? Raidboxes are an important affiliate link for Kevin Ohashi, but what about his integrity?
There will be no change in Kevin Ohashi’s methods. Kevin Ohashi is immune from criticism due to his dodgy relationships with many Facebook WordPress group admins and “experts”. The administrator of 25 WordPress groups on Facebook, Ivica Delic, has a personal interest in promoting Kevin Ohashi tests as the most objective. As well, Ivica Delic ensures that any critic of Kevin Ohashi within his WordPress groups on Facebook is cancelled. Why bother with the obvious flaws in the recommendations if you have such friends?
So, it is teamwork. Some “experts” will try to convince you to switch to a shared hosting plan on slow hardware, and the other will try to convince you that there is nothing faster and more secure (sic!) than unmanaged Vultr HF VPS on Cloudways. Both will collect affiliate fees by misleading and pushing an unsuspecting audience to switch to the wrong service provider. The perfect symbiosis and coordination in various WordPress groups on Facebook!
In most cases, users are surprised to learn that there is a third option – premium shared hosting providers with hardware that is as good or better than unmanaged VPS. With cPanel and LiteSpeed Enterprise, such premium providers offer you a much more secure and better web hosting experience. They may not have an army of affiliate warriors behind them, but that does not make their services any less attractive.
What else did we learn from our study? Scores on different instances of Vultr HF can vary widely. Vultr HF offers three different types of servers. The first type is a cone – it is only 4% faster than normal Vultr servers and 3% slower than DigitalOcean Premium Intel servers. The second type is 7% faster than DigitalOcean Premium, and the third type is 17% faster than DigitalOcean Premium. We tested seven Vultr HF servers and found that three of them were slowest type and three were fastest type. All are reported as Intel Skylake 3.7 GHz. Vultr HF plans may experience erratic movement between machines in the Data Center, which can affect your site’s performance. We compared LiteSpeed server providers with the slowest Vultr HF servers.

Additional PHP benchmarks

K6_loader_vs_CPU_score
Data from the K6 loader was normalized by the number of CPU cores in the plan. The main “blue” group shows a surprising correlation with the CPU performance score.
Finally, we would like to point out the perfect correlation between the results of the K6 load test and the CPU score derived from our study. There is no question that the companies of the Orange group perform the worst in the “Operations with data” sections of the WPBenchmarking plugin, “Math” and “String manipulations” sections of the WPPerformanceTester plugin, and “Calculations” section in the Benchmark.php test. Furthermore, testing with the  PHP simple benchmark script described below confirmed that the EPYC 7302 servers used by A2Hosting and NameHero actually achieve 66% of the performance of the EPYC 7502 servers used by Brixly. The K6 load test figures above is solid proof of this.
In the overall PHP benchmark tests, Xeon E-2136 used by EthernetServers also show a reduced score, dropping to the level of Skylake servers used by Guru. However, actual K6 load performance is still about 10% lower. We don’t have another E-2136 server to compare to, so it may be due to problems within the sysadmin department. Still, it is more likely to be a fundamental issue with this specific hardware.
In the K6 load test of uncached simplistic vanilla websites, AMD EPYC 7302 and Xeon E-2136 servers showed unexpected performance drops, suggesting that these servers are not always the best for serving complex dynamic websites with many PHP calls to the database. They are best suited to serving simple static brochure websites, where you will be using server processing power primarily for editing website content.
A few words about how to use a PHP benchmark script – it is available on Github. To measure these benchmarks, you need to copy four files – bench.php, common.inc, php5.inc, and php7.inc – to the website folder. If you want to run the PHP test, you must execute it via the browser https://yourdomain.com/bench.php?memory_limit=512 . The results will appear onscreen.
In the table below, we show PHP performance for two instances of Vultr HF. As we explained previously, Vultr data centres have several types of Intel hardware, which, in our tests, show very different speeds. We used a slower instance as a reference point for our main benchmarks. However, in this PHP benchmark, we present both the slowest and fastest instances explicitly. They are both Intel Xeon Skylake with a frequency of 3.7 GHz. However, they are entirely different classes of servers. The original benchmark shows up to a 20% difference, but the PHP benchmark identifies a variation of about 30%. That’s almost as much of a difference as between EPYC 7302 and EPYC 7502 servers. Once again, we see two processors that seem to belong to different generations and yet are branded as the same family. Please be aware that you can’t control what hardware is used on your account at Vultr. Even if you get a winning ticket and are served on the fastest instance, you can be moved to a slower instance later.
CPU_PHP_benchmark_normalised
Normalised results of the PHP simple benchmark script .
The PHP benchmark measures the speed of specific PHP functions and outputs it as a number of operations per second. Based on the measurements of a large number of servers in the group, we can determine a medium value for each PHP function separately. We calculate the deviations in percentages between individual server speed and a medium speed for a specific PHP function and record it in the table above. The sum of the values in each string should be equal to zero. The last row records those 22 metrics averaged across each server (across the column). This is the deviation of the PHP processing speed of this particular server from the “medium server” in the group. All of the values in the line still add up to zero. However, the medium value of the group is arbitrary and is determined by the selection of the servers. Therefore, we recalculate individual values to normalise them to the slow instance of Vultr HF, just as we did in our main benchmark section. The new normalised PHP metric is shown in the fifth row of the table.
Calculating the overall PHP metrics score is not an exact science. We should use different weights for different PHP functions to create a representative profile for the smart metric. However, this is beyond the scope of our study. Our goal was to establish a rough correlation between our simplistic CPU score, K6 load data, and our simplistic PHP benchmark. Moreover, we are glad to see that PHP benchmarks can help explain the knocking down of AMD EPYC 7302 and Xeon E-2136 shown by the K6 load tests.
The purpose of this exercise was also to confirm that Ryzen 9 5950x and E-2288G servers can significantly outperform Vultr HF servers. It might surprise readers, but the WordPress Hosting Facebook group is full of enthusiasts who take personal offence at a statement that Intel and AMD produce better servers than the one available at Vultr. We were taken aback by the worshipping of Vultr by VPS cargo cult followers. Vultr in our universe does not define the marketing and development strategies of Intel and AMD. Our PHP metrics show that Xeon E-2288G, E-2388G, and AMD Ryzen Zen 3 outperform Vultr’s fastest instance. The AMD EPYC 7502 server falls between the fastest and slowest Vultr instances. The new EPYC Zen 3 7543 servers should, however, surpass the fastest instance at Vultr.
Currently, you can find the fastest servers with enthusiastic small and medium shared hosting providers. Vultr, DigitalOcean, and many other larger shared hosting providers will be using slower servers. Hetzner does not offer Ryzen Zen 3 servers for VPS accounts, but shared hosting providers can get them as bare-metal servers. Any reasonable shared hosting provider will keep an eye on the CPU utilization of their servers to ensure that it stays below 50% on average. We bet that most VPS providers like Vultr and DO do not have an incentive to do the same for users of shared VPS accounts.
A semi-dedicated shared hosting account from many providers, including MechanicWeb and Brixly, can offer you more CPU and RAM resources than most expensive VPS plans. More importantly, they will implement security and malware protection at the server level – something that many VPS enthusiasts miss when using management panels. Choose a shared hosting provider that offers AMD EPYC 75×3 or Xeon E-2288G servers, and don’t be afraid to get your hands on the fastest Ryzen 9 5950x and Xeon E-2388G!

Case study: using PHP benchmark to identify overcrowded servers

Above, we introduced PHP benchmarks as a way to measure the performance of multiple tests. Such benchmarks can be successfully used to identify overloaded servers. It is important that you measure PHP benchmarks up to four times during a day and about four times during a night. It is possible that the average performance at night is much faster than the average performance during the day. You will see something like this if your server is overloaded. With overloaded servers, the results during the day are usually noisy. The deviation between maximum and minimum value can be more than 30% for each test.
E2288G-tests
 The PHP tests were run on E-2288G servers in Singapore at the Vultr data center (Stablepoint) and in Los Angeles (MechanicWeb).
The table above shows average results measured on the E-2288G servers at Stablepoint and MechanicWeb over daytime and nighttime. We collected data using PHP 7.4 with OPcache disabled. Both companies offer CloudLinux, cPanel, and LiteSpeed Enterprise. Therefore, similar results would be expected. Surprisingly, performance on the Stablepoint server during the daytime was two times slower. Their daytime benchmark data fluctuated by as much as 40% for each test. During the night, performance jumps by 50% and becomes less noisy. With MechanicWeb servers, benchmarks are much more stable, showing no difference between day and night metrics within error margins. PHP benchmarks fluctuate by about 3%.
There is no doubt about the conclusion. MechanicWeb keeps its production servers underutilised, while Stablepoint is not afraid to host shared sites on crowded servers.
Our analysis suggests that Stablepoint is running its servers at up to 100% logical CPU utilisation. This is highly unprofessional and rarely done by even the most outrageous hosting companies. The operating system creates two times more logical CPU cores than the number of physical cores. At a logical CPU utilisation of 100%, all physical CPU cores are multiplexing between two logical cores via Hyper-Threading. According to Intel, Hyper-Threading increases CPU efficiency by up to 30% by eliminating CPU idle time. As a result, each logical CPU core has access to 130%/2=65% of the CPU core’s performance. This is exactly what happens with the Stableopint server during the day. At night, the logical CPU utilisation drops below 50%, which means that all physical CPU cores are running a single logical core or are idle. As a result, the logical CPU performance is improved to 100% of the physical core. Since 65%*1.5=100%, this explains why the Stablepoint server is 1.5 times faster at night.

Next steps

Our main research article provides a quick overview of the best hosting options currently on the market. The following posts continue the research article.
  • Part 1. We give a detailed overview of the standard software options offered by all reasonable LightSpeed web hosting providers. Before you choose your hosting, you should read this chapter!
  • Part 2. You should read the detailed chapter explaining the cheating and fraudulent schemes of Cloudways and its affiliate community. We strongly recommend you to study it if you are still angry at us after reading our direct insults against Cloudways and their iconic services.
  • Part 3 (this post). In the third chapter, we explain the technical details of our benchmarking study. Read it if you want to quickly test the actual quality of hardware resources provided by your hosting provider. You’ll also get an objective look at their sysadmin team – can they fine-tune software settings to optimize database performance?
  • Part 4. The final chapter covers advanced hosting options for extremely busy websites. You should study this chapter if you have more than 100,000 visitors per month and are determined to switch to VPS hosting.
There are many hosting providers – each with its own advantages and disadvantages – but it can be hard to decide which is the best one. We have researched and sorted out the most appropriate hosting providers who offer all of the services discussed in this article’s first section. Our separate report has covered in detail some of the providers you will have to choose from. The obvious winner is MechanicWeb who provide their plans at the most affordable price using fastest and most optimised servers.
When websites attract heavy traffic, they overwhelm the provider’s hosting capacity. Adding a CDN service is an excellent way to get your website in front of more people and grow your audience. If you have a website with high traffic, then we recommend giving BunnyCDN a try. It will store a permanent cache of your website’s static files like CSS, JS, and images in its Geo-Replicated Perma-Cache servers. These servers will provide your files to the cache of 30 servers located around the world, shielding your origin from traffic. Your origin server will be left with the task of delivering HTML files to visitors, reducing the bandwidth load at your origin by a minimum of thirty times. BunnyCDN services cost you about $1 per month. It is a much better option than upgrading the level of your subscription with your original hosting provider.
If your website is hosted on a LiteSpeed Web server, you can reduce your workload by serving HTML pages through QUIC.cloud CDN. This CDN gives 10 GB of free traffic per month, leaving your origin server to only serve dynamic HTML files for your eCommerce clients. Please read our blog post “How to Choose Your Best CDN Provider” for more in-depth technical information about the services discussed above.
Web design agencies are essential for the success of any website. Selecting a proper hosting and content distribution network provider can help improve the page loading time but to succeed, you need a professional web design. Website design tools are increasingly commoditised, leading to an increase in the number of web designers with limited technical knowledge. Many of these designers cannot guarantee high-speed page loads for mobile devices or desktops. Our article “How to Choose a Web Design Agency” includes clear criteria for evaluating web design companies.
We’re one of the best web design agencies in Cambridge, and we specialise in speed optimisation. Our company has been in the graphic design industry for years, with over seventeen years of experience. We work hand in hand with our customers to understand their needs and create a strategy exclusively for them. We provide a comprehensive design process combined with an affordable web design package to all our clients. Please visit our website’s page about our web designs services for more information on what we offer.
Through our comprehensive solutions, we can help your business from logo designs to designing eCommerce stores. We provide a range of services, including technical SEO and relocation assistance to better web hosting. With our graphic design expertise, we are prepared to meet any other creative need that may arise.
Please contact us at svetlana.zh@webwhim.co.uk if you need any assistance, and we will be happy to help you.
Visit our WordPress research portal for more information. We spoke with plugin developers and hosting providers to compile six articles about how to become a better WordPress user. A knowledgeable web design agency will make all the difference in the speed of your site. You also need a set of plugins, tools, and a good hosting provider and a fast CDN to keep up with incoming traffic. If you don’t do these basics, your website will not pass the Page Experience signals introduced by Google.