Cloud / VPS Apache Performance Comparison

UPDATE #1: I received several comments about the testing methodology used in these benchmarks so I wanted to explain further. Each server instance at the Rackspace Cloud, Linode and Amazon are running the exact same Linux distribution (Debian 5.0 Lenny), kernel, architecture (x86_64) and version of Apache (2.2.9 w/ worker MPM).

UPDATE #2: After much concern was raised I went ahead and reran the benchmarks on the small instances after tuning a few settings on the Linode server.

echo 99999 > /proc/sys/net/ipv4/netfilter/ip_conntrack_max

Also, instead of running the tests from a consistent remote location I ran them on the local machines this way I could guarantee that network latency wouldn’t impact the test results in any way.

—–

In the past few weeks there has been a lot of talk about the differences between various Cloud and VPS providers (see my post), mainly revolving around performance. I decided to take some time last week to run some very straight forward and real world tests to gauge performance among the most popular providers: The Rackspace Cloud, Amazon EC2 and Linode.

For the benchmark I decided to use ApacheBench and Siege as *most* users on these platforms are serving some type of web content via Apache. Even if different web server software is used, it is unlikely that the results and trends would differ. As part of a series of posts, I will be running the same benchmarks against static content and dynamic content. This way we can be sure to look fairly at all aspects of each particular platform.

As mentioned, I decided to use ApacheBench and Siege as the testing platforms so I performed a default install of Apache 2.2 with the Worker MPM on each server instance. The server OS platform is Debian 5.0 Lenny on all servers.

These benchmarks are very easy to duplicate, so if you would like to test the results simply replicate the total connections or time and concurrency settings mentioned below.

I did not tweak any of the Apache settings on any of the server instances. I used the default settings of:

<IfModule mpm_worker_module>
StartServers                             2
MaxClients                              150
MinSpareThreads                  25
MaxSpareThreads                 75
ThreadsPerChild                    25
MaxRequestsPerChild          0
</IfModule>

I ran the same tests against each platform and since I was only hitting the default HTML page generated by the Apache install I decided to raise the concurrency limit up to something significant to generate some real load. Each benchmark was run three times over three different days and I took the best result from each set of tests.

ApacheBench: 100,000 Total Connections / 100 Concurrent Connections
Siege: 10 Minutes Under Siege / 50 Concurrent Connections

Let’s get to the results!

Small Size Instances

ApacheBench Results

Rackspace 256MB Cloud Server Linode 360MB Media Temple DV 512
Concurrency Level 100 100 100
Requests [#/sec] (average) 10,706.64 4,554.43 N/A
Time Per Request [ms] (mean) 9.340 21.957 N/A
Transfer Rate [Kbytes/sec] Received 3336.14 1,414.39 N/A

Siege Results

Rackspace 256MB Cloud Server Linode 360MB Media Temple DV 512
Transactions 1,587,997 1,165,412 N/A
Availability 100.00% 100.00% N/A
Response Time 0.04 secs 0.05 secs N/A
Transaction Rate 2,647.68 trans/sec 1,942.90   trans/sec N/A
Successful Transactions 1,587,997 1,165,412 N/A
Failed Transactions 0 0 N/A
Longest Transaction 0.42 0.61 N/A
Shortest Transaction 0.00 0.00 N/A

Review

The Rackspace Cloud Server instance, with only 256 MB of RAM, performed well under the stress as did the Linode 360 server, but there is still a significant difference in the overall performance. This could be due to host server utilization, but according to the Linode control panel the host machine my instance was running on is “low”.

In the Siege results we get some similar insights. The 256 MB Rackspace Cloud Server processed significantly more transactions than the Linode server. Importantly, the transaction rate was also much higher.

There is always a lot of discussion around the Linode 360 server instance being a better value than the 256 MB Rackspace Cloud Server, but the tests show otherwise. In my opinion this is most likely due to Rackspace Cloud’s lower utilization of host hardware and a more robust network.

Large Size Instances

Now lets take a look at the larger instances. I wanted to work Amazon’s EC2 platform into these performance tests and their large instance is most comparable to the Cloud Server and Linode instances of the 8GB variety.

ApacheBench Results

Rackspace 8192MB Cloud Server Linode 8640MB Amazon EC2 Large
Concurrency Level 100 100 100
Requests [#/sec] (average) 10,729.61 9,109.91 6,782.93
Time Per Request [ms] (average) 9.320 10.977 14.743
Transfer Rate [Kbytes/sec] Received 3,363.62 2,829.17 2,113.04

Siege Results

Rackspace 8192MB Cloud Server Linode 8640MB Amazon EC2 Large
Transactions 1,564,152 1,636,652 891,956
Availability 100.00% 100.00% 100.00%
Response Time 0.04 secs 0.04 secs 0.07   secs
Transaction Rate 2,606.79 trans/sec 2,727.34   trans/sec 1,485.80   trans/sec
Successful Transactions 1,564,152 1,636,652 891,956
Failed Transactions 0 0 0
Longest Transaction 0.49 0.22 0.65
Shortest Transaction 0.00 0.00 0.00

Review

As we saw with the smaller instances in the ApacheBench test, the Cloud Server with fewer resources is outpacing both the Linode server and the Large EC2 instance. But in the Siege benchmark we see the Linode instance take the crown for total transaction rate. I would note the larger memory allocation on the Linode server instance might have something to do with these results so be sure to compare price for your particular application.

Amazon’s Large EC2 instance was the slowest of the bunch during the ApacheBench testing but did complete all of the transactions thrown at it during the Siege benchmark.

Final Thoughts

The tests show some significant performance benefits when running on the Rackspace Cloud Servers platform on the low end and similar performance to larger Linode and Amazon instances on the high end even though the comparable Cloud Server had much less memory and therefore lower cost. In my opinion, I believe lower hardware utilization and better host server hardware contribute to these results.

Apache settings can definitely be tweaked to increase performance on any of these platforms; this was just a baseline test using the defaults provided. The next round of tests will focus on serving a dynamic web applications.

If you have any thoughts, questions or comments about the benchmarks performed here please leave a comment below. I’d love to hear from you!

%d bloggers like this: