Australian Shared Hosting Benchmark 2018
Published and retracted 2018-11-11, re-published 2018-11-28
The purpose of this project was to figure out the following:
Who, of the most popular Australian web hosts, provides the fastest cPanel hosting at the $20/month price point?
It follows up on a similar question I asked in 2017: Who is the biggest cPanel web host in Australia?
It was also important for the benchmark to clearly justify all of its decisions and to provide a way for others to reproduce the results on the same or similar kinds of services. To that end, every part of this benchmark is open source, including the tools and raw data, so you can run any or all parts of it yourself.
A caution: Remember, these results only reflect around a week (~4 daily runs per provider) of testing, Providers update their services constantly, and all benchmarks in existence are flawed. Consider that when choosing how seriously to take these results in your decision-making.
Providers and Services
The specific cPanel services that were benchmarked were selected according to the Methodology, which can be approximately be summed up as: take the ~10 most popular web hosts on Whirlpool and choose their ~$20/mo plan.
These are then benchmarked against a VPS service at the same price, configured in a performance-optimal way.
|Provider||Plan Name||Cost/Month||Hostname||CloudLinux||Webserver||PHP Version||Database Version||Kernel|
|Linux VPS (2 CPU, 2GB RAM)||$20.00||nerve-idiom||No||nginx/1.15.5||7.1.23||5.5.5-10.2.18-MariaDB||3.10.0-862.14.4.el7.x86_64|
|Digital Pacific||Business Standard||$21.90||vmcp44.digitalpacific.com.au||Yes||LiteSpeed||7.1.24||5.6.41-cll-lve||2.6.32-896.16.1.lve1.4.51.el6.x86_64|
|Melbourne IT||cPanel® Starter Hosting||$14.94||cp374.ezyreg.com||Yes||Apache||7.0.32||5.6.41||2.6.32-673.8.1.lve220.127.116.11.el6.x86_64|
|Micron21||cPanel - Custom Plan (200% CPU, 2GB RAM, 10GB Storage)||$23.10||cp-kil-m-009.micron21.com||Yes||LiteSpeed||7.1.24||5.6.41-log||3.10.0-714.10.2.lve1.5.9.el7.x86_64|
|Net Virtue||Business - Basic||$19.95||b9.cpcloud.com.au||Yes||LiteSpeed||7.1.24||5.5.5-10.2.18-MariaDB-cll-lve||3.10.0-714.10.2.lve18.104.22.168.el7.x86_64|
|Panthur||Business Hosting - Bronze||$20.00||vmcp22.web-servers.com.au||Yes||LiteSpeed||7.1.24||5.5.5-10.0.36-MariaDB-cll-lve||3.10.0-714.10.2.lve1.5.12.el7.x86_64|
|Zuver||Web Hosting w/ Power Pack||$6.00||flash.zuver.net.au||Yes||LiteSpeed||7.1.24||5.7.24||2.6.32-896.16.1.lve1.4.54.el6.x86_64|
Synergy Wholesale, TPP Wholesale, Crazy Domains, and Net Registry were all disqualified, bringing the field down to 8 providers.
Overall Winner: Netorigin
The overall winner is the Perth-based Netorigin's "Elite" cPanel service. It performed outstandingly on all benchmarks, exceeding the VPS in a number of cases. 🎉
and on the eastern coast ...
Since, as we all know, nobody lives in Perth, it wouldn't be fun unless we also crowned an east-coast winner. That glory belongs to VentraIP, who came close enough to Netorigin on the WP Overall Score, that you could argue for a draw.
|Provider||WP Overall Score||WP P90 1/sec||WP P90 5/sec||WP P90 10/sec||DB Insert||DB Queries||Prime Sieve||PHP.net #1||PHP.net #2||IO Random||IO Open||IO Seq. Write|
2.40x slower overall
|27.76ms 1.20x slower||91.65ms 2.60x slower||134.27ms 2.85x slower||21.07s 1.76x faster||76 1.46x faster||11250 1.36x faster||298.00ms 1.40x faster||2031.00ms 1.44x faster||95951 4.57x slower||35.84ms 1.50x faster||64.76ms 2.93x faster|
2.79x slower overall
|56.58ms 2.45x slower||107.39ms 3.05x slower||129.98ms 2.76x slower||34.15s 1.08x faster||47 1.11x slower||6809 1.21x slower||504.00ms 1.21x slower||3207.00ms 1.10x slower||99867 4.39x slower||80.42ms 1.50x slower||120.03ms 1.58x faster|
3.42x slower overall
|95.23ms 4.12x slower||105.79ms 3.01x slower||159.74ms 3.39x slower||71.27s 1.93x slower||31 1.68x slower||6210 1.33x slower||915.00ms 2.19x slower||4331.00ms 1.48x slower||62396 7.03x slower||652.50ms 12.17x slower||102.88ms 1.85x faster|
4.38x slower overall
|67.39ms 2.91x slower||175.87ms 5.00x slower||218.49ms 4.63x slower||54.80s 1.48x slower||41 1.27x slower||6494 1.27x slower||611.00ms 1.47x slower||3728.00ms 1.28x slower||3768 116.33x slower||244.68ms 4.56x slower||31380.02ms 165.27x slower|
4.55x slower overall
|78.27ms 3.38x slower||221.44ms 6.29x slower||180.86ms 3.83x slower||47.50s 1.28x slower||39 1.33x slower||6502 1.27x slower||511.00ms 1.23x slower||4097.00ms 1.40x slower||345645 1.27x slower||91.32ms 1.70x slower||86.75ms 2.19x faster|
6.28x slower overall
|99.97ms 4.32x slower||196.48ms 5.58x slower||366.08ms 7.76x slower||82.48s 2.23x slower||34 1.53x slower||4844 1.70x slower||926.00ms 2.22x slower||5442.00ms 1.86x slower||16685 26.27x slower||114.71ms 2.14x slower||2554.74ms 13.46x slower|
6.79x slower overall
|107.07ms 4.63x slower||212.61ms 6.04x slower||396.54ms 8.41x slower||83.61s 2.26x slower||35 1.49x slower||5488 1.50x slower||683.00ms 1.64x slower||4582.00ms 1.57x slower||15774 27.79x slower||219.78ms 4.10x slower||211.20ms 1.11x slower|
890.99x slower overall
|847.87ms 36.64x slower||42600.00ms 1210.23x slower||50560.00ms 1071.87x slower||58.65s 1.59x slower||46 1.13x slower||2007 4.11x slower||2408.00ms 5.77x slower||14011.00ms 4.80x slower||3841 114.12x slower||161.16ms 3.01x slower||30478.16ms 160.52x slower|
Visualized (based on WP Overall Score)
The selection of providers is a task that could be strongly affected by bias, so I endeavored to give it some semblance of being data-driven.
The ideal scenario was to scrape all of the threads posted in during 2018 in the Whirlpool Web Hosting forum, apply named-entity recognition on the posts and then perform a frequency count for each named-entity (web host).
Unfortunately I did not quite reach that scenario. While the scraping of the forum was easy to perform (and with permission), the state-of-the-art for NER requires a significant amount of training to be at all effective. e.g. "net registry" is very hard to pick up as an NE. In the end, I (very unglamorously) used my eyeballs to identify the top 17 web hosts from the ~16k posts.
From there, permissive RegExps were written for those 17 and used to generate the frequencies.
One significant mistake that I made was counting all posts whether they were made in 2018 or not (so threads that go for many years had all posts counted). A second mistake was that posts from another forum ("Server Management") accidentally leaked into the dataset during the scrape, but I suspect its impact was negligible.
You can look in the source code repository for all of the raw data and the actual frequencies.
A meaningful measure of "performance" in shared web hosting should ideally reflect the speed at which humans experience the web applications that they run on their cPanel service. As such, we will run only pure-PHP benchmarks which are initiated over HTTP requests. This rules out usual tooling like sysbench, FIO, etc. Another reason for using pure PHP is that a number of hosts have disabled system/exec and it's not clear whether it's possible to fork an arbitrary ELF executable (e.g. using cgi-bin) across every host or not. Since it it has a chance of torpedoing the entire methodology, I chose to avoid the risk.
In benchmarking, it's unavoidable to rely on microbenchmarks. Even for "realistic" tests, they will not truthfully predict a real-world workload. However, this is a comparative study. We are not gunning for accuracy in the absolute numbers reported by the benchmarks, just for accuracy in the comparative numbers. For this reason, the results are shown as a percentage of the performance of a baseline environment (a fast VPS).
The baseline environment is an unmanaged Linux VPS from Binary Lane, at the same cost target as the shared hosting plans. It is configured in a performance-maximizing way (nginx+PHP-FPM), with the aim of providing some guidance as to "how much slower/faster" a shared service is, compared to an optimum (though not necessarily the optimum ).
Each benchmark is also repeated 25-30 across a week time interval, in order to account for transient "noisy neighbor" effects.
The pure PHP benchmarks are described below.
This is the "realistic" benchmark that primarily informs the overall result.
It involves installing an entirely vanilla WordPress installation and performing a constant-rate, concurrent HTTP siege against it using wrk2, the "coordinated-omission" fork of wrk. I have further forked wrk2 at the ausdomainledger/wrk2 multi-ip branch in order to round-robin requests between IPv4 and IPv6 addresses, when available. The reason for this is that Zuver and VentraIP have extremely aggressive CSF firewall rules configured, which causes the 10/sec concurrency level to fail. Using multiple source IPs permits us to evade the firewall.
The aim of the benchmark is to record the latency percentiles for the site's transaction time at fixed concurrency levels (1/sec, 5/sec, 10/sec). These should approximately reflect the usability of the website when it is being browsed by only the webmaster, by a few users, and a few more users, respectively. Initially 10/sec was 25/sec, but it became clear during benchmarking that all but one of the providers fell flat on their faces under such a workload.
Due to shared hosting being a multi-tenant environment, we will focus mainly on the P90 result, as anything higher is more or less just noise. However, you may browse the raw data for a complete latency distribution for each provider.
The "siege" is performed from a dedicated server hosted on Servers Australia's Sydney network, and is within a millisecond latency of every benchmarked website. The first exception is Micron21, who are hosted in Melbourne, who are sieged from a separate VPS node hosted in Micron21's Kilsyth DC. The second exception is Net Origin, who are hosted in Perth, who are sieged from a separate VPS node hosted by RansomIT in Perth. In both cases, extra care was taken to ensure that the different environment did not introduce extra advantage or penalty.
This is a second "realistic" benchmark which focuses on the ability of the user to put the database server to work. It is significant to include because each host may or may not be running CloudLinux's MySQL governor.
The benchmark is based upon the first 500k records in the ASIC Business Names dataset, with a full-text MySQL search index on the business name column. The benchmark itself reports on the throughput of INSERTs and full-text queries.
3. Microbenchmark: CPU - Prime Sieve
This is a better CPU microbenchmark than #4 and #5, basically because it runs longer. It is my own port of Sieve of Eratosthenes used by sysbench to measure single-threaded CPU performance.
4. Microbenchmark: PHP.net's bench.php
This is one of the two official microbenchmarks used by PHP to report on performance gains and regressions between PHP versions. However, it runs very quickly and is not very good at measuring any meaningful workload.
5. Microbenchmark: PHP.net's micro_bench.php
6. Microbenchmark: Disk IO
Disk IO is a terribly important subsystem when measuring performance, but it is more or less impossible to accurately measure from a pure PHP benchmark. For this reason, it is heavily discounted in the overall result.
Nonetheless, an attempt is given to measure this, for curiosity, if no other reason. The benchmark essentially measures mass random read/write/open on inodes characteristic of the kind you would find in a shared hosting web application (e.g. WordPress).
bUt iS iT sTaTisTiCaLlY sIgNiFiCanT?
Directly comparing samples (i.e. percentiles) across datasets is potentially problematic. I haven't proven that the measured differences did not occur purely by chance.
By running the benchmark suites 25-30 times across different days and for significant durations, I have tried to provide an accurate basis for comparison, but my good feeling about it is purely intuitive rather than benefiting from rigorous stats analysis.
Ranking the Results
For the overall rankings, the sum of the median of each of 1/sec, 5/sec and 10/sec P90s is taken.
It might be surprising to ignore the other measurements entirely, but coming up with a composite/summary score based on some kind of factor or principal component analysis proved to be too unwieldly for me to do correctly, and in a reasonable time.
This ranking method is also the most direct measure for the idea of "performance" that most people think about in the context of shared hosting, in an environment that is more or less homogenous towards and optimized for WordPress workloads.
In any case, you are free to sort the results by whatever measurement you want.
Disclosures and Contact
Of all of these providers, I am affiliated only with the baseline benchmark provider, Binary Lane, as an existing customer. Their selection as the benchmark provider was made on the basis of personal preference and the fact that they are not eligible as a shared hosting provider.
No providers were informed about being benchmarked at any time. Care was taken not to reveal the purpose of the accounts.
All providers were paid the full amount for their services using only my personal funds.
Services were used only for the purpose of this benchmark, and then cancelled. I received an automatic refund from VentraIP & Zuver at cancellation time, apparently in accordance with their 45 day money-back guarantee.
Feedback can be provided by 📧 to _<at>ausdomainledger.net. I am happy to post factual corrections or highlight mistakes and flaws. However, no existing or additional benchmarks will be re-run or have their results updated.
All product names, logos, and brands are property of their respective owners. All company, product and service names used in this website are for identification purposes only. Use of these names, logos, and brands does not imply endorsement.
 Popularity is calculated by the number of mentions on the Whirlpool Web Hosting forum in 2018. See Methodology for more info.
 $20/month was selected as an arbitrary, comfortable price-point, that feels like the mean for what is offered by hosts and chosen by customers, but isn't backed by data.
 Caching plugins are entirely orthogonal to the purpose of this benchmark. We are trying to perform a comparative study of the "raw power" of each cPanel service. To that end, any caching plugins that are snuck in by the server (like Litespeed Cache) are forcibly removed.
 PHP does not expose a sufficient set of syscalls to make it possible to control whether a write is actually applied to disk. At most, disk writes will result in dirty pages in the kernel, so the benchmark results will ultimately depend on the state of the kernel page cache at any point in time (in other words, completely out of our control).
 The selection of Binary Lane as the benchmark baseline is not any kind of endorsement or statement about its worthiness as a web host. It is purely a question of convenience and independence from the benchmarked providers.
 Synergy Wholesale and TPP Wholesale were disqualified for not having any retail (not requiring pre-approval/contract signing) services available. Crazy Domains were disqualified for requiring annual+ payment terms, which does not fit the spirit of the problem statement. Net Registry were disqualified for unironically asking me to fax them my drivers licence, but it is unlikely that their results would differ from that of Melbourne IT by much.