Blog Posts

I just posted a new whitepaper that I have been working on for the last month or two. It was a great opportunity to test out how well Hyper-V performs with a large number of virtual machines (VMs) on a richly configured R900. You'll have to check out the whitepaper for the complete details. Here I'm going to give you a little bit of the background into how the paper was done.

I found out in early summer that we would be announcing and shipping some new 6-core Intel Xeon processors in our PowerEdge R900 server (4-socket, 4U). I was being asked what I thought would be a good way to show how well it performed for virtualization and recommended that we use Hyper-V and the Dell DVD store in a test similar to some tests that I had done in the past with ESX. At the time, Hyper-V was still just a release candidate, and I was told that I wouldn't get the processors until early August. Both of these things changed-Hyper-V went officially final with an RTM build (which was great), and I ended up not getting any 6-core processors to test with until late August (which was bad because I had only one week to complete testing at that point).

I switched over to EqualLogic for the storage in this test. Previous tests had mostly been with Dell | EMC CX series storage, but I wanted to get a chance to run a heavy load against the EqualLogic iSCSI storage. It was fun to set up and test with the new arrays. I thought that setup was much easier, and I learned about setting up MPIO with the iSCSI initiator on Windows 2008.

Showing that the R900 with 24 cores was able to support 40 VMs, achieving a total of 74K DVD Store Orders per minute, wouldn't mean anything unless it was compared to something else. Using an R900 and an HP DL585 as quad-core comparison points, the testing showed an advantage in not only performance, but performance per watt as well.

So check out the full paper, and let me know if you have any questions.

In response to a few questions that came up, I put together a video that shows how to do a cool trick with Microsoft Excel to create graphs of MD3000i array level performance data. I posted an entry here last week and also put up a wiki page about how you can use the smCLI command line tool to get array level performance data (Individual storage processors, array totals, and individual virtual disks) from an MD3000i. The output of this command is a csv (comma separated values) text file. I included on the wiki page a nice graph of the performance based on the data from this output file. What I didn't go into was how to create such a graph.

So now you have this great little video that will show you one way that you can use Excel to create a graph based on the data in the file. I don't claim to be a know-it-all when it comes to Excel, but the way that I do it in the video works. I would love to hear from anybody who knows of other (possibly better) ways to do it.

Todd

A few months ago Dan Hambrick from our System Performance Analysis team had completed and published a performance report of what I thought to be a very complete and thorough look at the PERC 6 (PowerEdge RAID Controller). It posted to our humble site and I did a blog entry to highlight how cool I thought it was.

About a month later Dan comes back to my cube and tells me that he is getting all kinds of requests to do more testing. Of course any new factor could grow his already very large matrix and double the amount of testing. He ends up figuring out how to do a round of testing with the new MD1120 array and get the data into a form that is understandable in a relatively small number of graphs.

The new paper was posted just a few days ago and the amount of data that is behind it is staggering. The numbers behind the graphs are all in the Appendix if you are brave enough to take a look, but I would highly recommend that you stick with the really cool graphs.

Todd

In the past I was somewhat biased towards 2-socket servers for virtualization - which was due to the results of some testing we did. I had of course realized over the past year or so that things had changed and the 4-socket servers were now more competitive. So when I ran the series of chats on Selecting a Server for Virtualization, I decided that it was a good opportunity to re-run some of those exact same tests with the brand new R900 that had arrived in the lab. I posted the results on a TechCenter page so you can see the full results, but the short summary is that the R900 stacks up as more efficient than the 2950 we tested in the previous paper. The R900 showed 10 to 23 percent better performance per watt than the older 2950. The question that remains is how would the R900 compare to a current generation 2950?

Todd

New VMmark benchmark results highlight that the best performance for a single server is a 4-socket (or 16 cores with Quad-Core processors). This means that the R900 and R905 are at the top of the list, with the Intel based R900 slightly ahead of the AMD based R905. So if you absolutely need the most performance possible from a single server - then this is the way to go.

A little bit of analysis could lead you to believe that the 2-socket (or 8 core) servers are actually better performing. The VMmark score for the 4-socket R900 is 14.23 with 10 tiles, but the VMark score for the 2-socket 2950 III is 8.47 with 6 tiles. So on a per socket basis, the two-socket 2950 III is actually providing more performance. The same holds true for the two-socket M600 blade and two-socket R805 2U server.

There is much more to choosing a server than the results of a single benchmark, but I think that these results are fair barometers of performance. In general performance does not scale in a linear fashion when moving from 2-socket servers to 4-socket servers, and this seems to hold true with virtualization.

Todd