Running databases in virtualized environments

I have long believed that databases can be successfully deployed in virtual machines. Among other things, that is one of the central ideas behind ParElastic, a start-up I helped launch earlier this year. Many companies (Amazon, Rackspace, Microsoft, for example) offer you hosted databases in the cloud!

But yesterday I read this post in RWW. This article talks about a report published by Principled Technologies in July 2011, a report commissioned by Intel, that

tested 12 database applications simultaneously – and all delivered strong and consistent performance. How strong? Read the case study, examine the results and testing methodology, and see for yourself.

Unfortunately, I believe that discerning readers of this report are more likely to question the conclusion(s) based on the methodology. What do you think?


A Summary of the Principled Technologies Report

In a nutshell, this report seeks to make the case that industry standard servers with virtualization can in fact deliver the performance required to run business critical database applications.

It attempts to do so by running Vware vSphere 5.0 on the newest four socket Intel Xeon E7-4870 based server and hosting 12 database applications each of which has an 80GB database in its own virtual machine. The Intel Xeon E7-4870 server is a 10 core processor with two hardware threads per core. It was clocked at 2.4GHz and 1TB of RAM (64 modules each of which had 16GB). The storage in this server was 2 disks, each of which was 146GB in size (10k SAS). In addition, an EMC Clarriion Fibre Channel SAN with some disks configured in RAID0. In total they configured 6 LUN’s each of which was 1066GB (over a TB each). They VM’s ran Windows Server 2008 R2, and SQL Server 2008 R2.

The report claims that the test that was performed was “Benchmark Factory’s TPC-H like workload”. Appendix B somewhat (IMHO) misleadingly calls this “Benchmark Factory TPC-H score”.

The result is that these twelve VM’s running against an 80GB database were able to consistently process in excess of 10,000 queries per hour each.

A comparison is made to the Netezza whitepaper that claims that the TwinFin data warehouse appliance running the “Nationwide Financial Services” workload was able to process around 2,500 queries per hour and a maximum of 10,000 queries per hour.

The report leaves the reader to believe that since the 12 VM’s in the test ran consistently more than 10,000 queries per hour, business critical applications can in fact be deployed in virtualized environments and deliver good performance.

The report concludes therefore that business critical applications can be run on virtualized platforms, deliver good performance, and reduce cost.


My opinion

While I entirely believe that virtualized database servers can produce very good performance, and while I entirely agree with the conclusion that was reached, I don’t believe that this whitepaper makes even a modestly credible case.

I ask you to consider this question, “Is the comparison with Netezza running 2,500 queries per hour legitimate?”

Without digging too far, I found that the Netezza whitepaper talks of a data warehouse with “more than 4.5TB of data”, 10 million database changes per day, 50 concurrent users at peak time and 10-15 on an average. 2,500 qph with a peak of 10k qph at month end, 99.5% completing in under one minute.

Based on the information disclosed, this comparison does not appear to be valid. Note well that I am not saying that this comparison is invalid, rather that the case has not been made sufficiently to justify it.

An important reason for my skepticism is that when processing database operations like joins between two tables, doubling the data volume quadruples the amount of computation that may be required. If you are performing three table joins, doubling the data increases the computation involved may be as much as eight times. This is the very essence of the scalability challenge with databases!

I get an inkling that this may not be a valid comparison when we look at Appendix B that states that the total test time was under 750 seconds in all cases.

This feeling is compounded when I don’t see how many concurrent queries are run against each database. Single user database performance is a whole lot better and more predictable than multi-user performance. The Netezza paper specifically talks about the multi-user concurrency performance not the single-user performance.

Reading very carefully, I did find a mention that a single server running 12 VM’s hosted the client(s) for the benchmark. Since ~15k queries were completed in under 750s, we can say that each query lasted about 0.05s. Now, those are really really short queries. Impressive but not what I would generally consider to be in the kinds of workloads that one would expect Netezza to be deployed. The Netezza report does clearly state that 99.5% completed in under one minute, which leads me to conclude that the queries being run in the subject benchmark are at least two orders of magnitude away!

Conclusion

Virtualized environments like Amazon EC2, Rackspace, Microsoft Azure, and VMWare are perfectly capable of running databases and database applications.One need only look at Amazon RDS (now with MySQL and Oracle), database.com, SQL Azure, and offerings like that to realize that this is in fact the case!

However, this report fails to make a compelling case for this. By making a comparison to a different whitepaper and simply relating the results to the “queries per hour” in the other paper causes me to question the  methodology. Once readers question the method(s) used to reach a conclusion, they are likely to question the conclusion itself.

Therefore, I don’t believe that this report achieves what it set out to do.


References

You can get a copy of the white paper here, a link to scribd, or here, a link to the PDF on RWW.

This case study references a Netezza whitepaper on concurrency, which you can get here. The Netezza whitepaper is “CONCURRENCY & WORKLOAD MANAGEMENT IN NETEZZA”, and prepared by Winter Corp and sponsored by Netezza.

I have also archived copies of the two documents here and here.

A link to the TPC-H benchmark can be found on the TPC web site here.

Disclosure

In the interest of full disclosure, in the past I was an employee of Netezza, a company that is referenced in this report.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s