Revision 5 posted to TechCenter Chats - Wiki by Lance Boley on 11/15/2011 8:38:59 AM
| DELL-ToddM | We're waiting a few minutes for people to join before starting |
| slbrown | Hello Todd. |
| DELL-ToddM | Any specific questions / areas you want to cover today? |
| Dell-RogerF joined. | |
| Dell-Darren joined. | |
| DELL-ToddM | Welcome Roger and Darren |
| Dell-Darren left. | |
| Dell-RogerF | Afternoon |
| DELL-ToddM | Just giving everybody a few minutes to join up |
| jampmcse joined. | |
| jampmcse left. | |
| DELL-ToddM | Let's go ahead and start the chat |
| DELL-ToddM | Today's topic is Virtualizing Sql Server |
| DELL-ToddM | Specifically the work that I have done with Sql Server has been on Vmware Esx Server |
| DELL-ToddM | Most recently (yesterday) I published a whitepaper with a performance characterization of Sql Server 2005 on Esx Server 3.5 |
| Dell-MarcFarley joined. | |
| DELL-ScottH | hey Marc |
| Dell-MarcFarley | what's up? |
| Dell-Darren joined. | |
| DELL-ToddM | There is a home page for the whitepaper on the site at http://www.delltechcenter.com/page/performance+of+sql+server+2005+on+esx+server+3.5 |
| DELL-ToddM | Hey Marc - just getting started |
| DELL-ToddM | In addition to the home page for the whitepaper, there is also a landing page for Sql Server on Vmware at http://www.delltechcenter.com/page/sql+server+on+vmware |
| DELL-ToddM | The paper really covers two sets of tests with 64-bit Sql Server 2005 Vms |
| Dell-Darren left. | |
| DELL-ToddM | The first set of tests measures the performance scaling as additional vcpus are added to a single Vm |
| Dell-Darren joined. | |
| DELL-ToddM | And how performance scales with multiple Vms |
| DELL-ToddM | And the best part - used lots of graphs with mulitple colors to illustrate the results! |
| DELL-ToddM | The overall finding for scaling within a single Vm was that performance scales about as well as a physical server would scale as you add vcpus or processors |
| DELL-ToddM | In other words - you don't get linear scalablity, but you do a healthy increase in performance |
| Dell-Darren left. | |
| Dell-Darren joined. | |
| DELL-ToddM | The reason for this test was that in the past, with previous versions of Esx, there was some concern about the performance of multi vcpu Vms |
| DELL-ToddM | In my testing, as reported in the whitepaper, the performance of the 2 vcpu and 4 vcpu Vm was great. |
| DELL-ScottH | which processors did you use in the server ? |
| Dell-Darren left. | |
| Dell-Darren joined. | |
| DELL-ToddM | The physical server used in the testing was a 2950 with quad-core 3.16 Ghz processors |
| DELL-ToddM | So there were 8 cores in the server, and 32 Gb of Ram |
| DELL-ToddM | Basically a max configuration |
| DELL-ToddM | In the single Vm testing the full resources of the server weren't really stressed - as a single Vm can only be assigned up to 4vcpus (which amounts to 4 cores) |
| DELL-ToddM | So to put the full power of the server to the test I did some multi Vm testing next |
| DELL-ToddM | Using the results from the single Vm test, a baseline of performance was established for the 1vcpu, 2vcpu, and 4vcpu Sql Server Vms |
| DELL-ToddM | I then ran tests with 2,3, and 4 of each of these to see how performance scaled across all the Vms |
| DELL-ToddM | The results showed that Esx was able to evenly distribute server resources across the Sql Server Vms - providing good performance scaling |
| DELL-ToddM | Each Sql Server Vm was able to achieve nearly the same amount of transactions per minute as the baseline - even though multiple Vms were running at the same time |
| DELL-ToddM | Up to a point |
| DELL-ScottH | is it possible to compare these results to running on a physical server ? .. perhaps if one was considering moving from physical to virtual to get a rough guestiumate on perf gains/loses |
| DELL-ToddM | Once the server processors became overcommitted (or over subscribed) performance scaled no further |
| DELL-ScottH | and where did you find bottlenecks in your testing ? |
| DELL-ToddM | The maximum number of transactions across all four Vms was limited to just over 50,000 Orders Per Minute - which was bottlenecked by the processors |
| DELL-ToddM | Each Sql Server Vm had a 100 Gb Dell Dvd Store database - with 20 dedicated datadisks and 2 dedicated log disks |
| DELL-ToddM | So everytime I added a Vm into the testing, it came with it's own disks |
| Dell-RogerF left. | |
| DELL-ToddM | A key point about running Sql Server in a Vm- |
| DELL-ToddM | You still have to size the storage correctly |
| DELL-ToddM | just like a physical server |
| Dell-Darren | Todd you may have covered this but what type of workload were you running? Oltp, Olap, Bi, etc.? |
| DELL-ToddM | Oltp - The Dell Dvd Store is a test app that Dave Jaffe and I developed and opensourced |
| DELL-ToddM | It is an online Dvd Store that allows you to logon, browse for discs, and purchase them |
| DELL-ScottH | I'm looking at your charts and the 4VM 2vcpu shows ~55K Opm and the 4VM 4vcpu shows ~50K Opm .... I can understand 4x4 and 16 cores being maxed out ... but did you try scaling beyond that in the 2vcpu tests ? .... maybe a 5th or 6th 2vcpu test ? |
| DELL-ToddM | It includes most standard Db stuff - primary and forgein keys, indexes, full-text indexes, and multi-phase transactional |
| DELL-ToddM | I ran out of disks with the 2vcpu Vms, so could not add more Vms to test |
| DELL-ScottH | ahh, makes sense, maybe Marc can help us out with that |
| DELL-ToddM | I was also out of Ram on the server - 8GB per Vm (actually had to drop back to 7.5 Gb for the 4 Vm tests) |
| DELL-ScottH | do you think it would continue to scale until max cores ? |
| DELL-ScottH | which parts on this Sql test bottleneck first ? .. Ram, disk, Cpu |
| DELL-ToddM | The Dvd Store will usually max out disk first, then Cpu. So you have to make sure that you have enough disks for the processors that you are using |
| DELL-ToddM | During initial phase of testing, perf monitoring showed that 20 data disks was enough to max out 4vcpu Vm |
| DELL-ToddM | I verified this by using 30 disks and getting no additional performance gain |
| slbrown | what type of data source where you connected to? |
| DELL-ToddM | The storage was a Cx3-80 fibre channel array |
| DELL-ToddM | Specific disk layout was 2 10 Disk Raid 1/0 (to make 20 data disks) and a Raid 1 for logs |
| slbrown | you may want to add that information to the test configuration as it provides some insight into what the backend storage was for the test |
| DELL-ToddM | It is in the full version of the whitepaper - the Pdf that is linked at the top |
| DELL-ToddM | The wiki page has the highlights |
| DELL-ToddM | I will add the storage configuration table to the wiki page - thanks for the suggestion |
| DELL-ToddM | Anybody else have any performance with Server stories they would like to tell |
| DELL-ToddM | Sql Server - that is |
| DELL-ToddM | The second half of the paper deals with Vmotion of Sql Server Vms |
| DELL-ToddM | One of the reasons to virtualize Sql Server is to be able to take advantage of Vmotion and the features that it enables |
| DELL-ToddM | So to me, it makes alot of sense to also measure the performance or impact of Vmotion with Sql Server Vms |
| DELL-ToddM | In order to test this, I created a test that would move a Vm using Vmotion back repeatedly over a 90 minute period |
| DELL-ToddM | While under moderate load |
| DELL-ToddM | In order to report the results I looked at a few things |
| DELL-ToddM | First was to make sure that Vmotion did not break Sql Server or cause lost transactions |
| DELL-ToddM | During all the Vmotion tests that I did for the paper, there were no errors in the event log |
| DELL-ToddM | Sql Server continued to process transactions, |
| DELL-ToddM | and no errors were reported from the client side |
| DELL-ToddM | So there were no problems found during testing - which is really important |
| DELL-ToddM | The time for the Vmotion to complete ( from the time the command was issued until the Vm was on the new server) did increase |
| DELL-ToddM | as the "size" of the Vm increased |
| DELL-ToddM | But the increase was not extreme |
| DELL-ToddM | The small 1vcpu Vm moved in 1min 49 sec |
| DELL-ToddM | The Large 4vcpu Vm moved in 3 min 6 sec |
| DELL-ToddM | There was a measurable difference in overall Orders Per Minute, but it was very small |
| DELL-ToddM | The most visible indicator of the Vmotion events was in the Cpu utilization of the Vm itself. |
| DELL-ToddM | There was a very brief spike during each Vmotion |
| DELL-ToddM | This was due to the brief queuing that would occur during the 1 or 2 that the Vm was actually moving from one physical server to another |
| DELL-ToddM | So, the net was that Vmotion works good with Sql Server Vms |
| DELL-ToddM | The funny thing is that this Vmotion test is very similar to the testing that I did in my first whitepaper on Vmware |
| DELL-ToddM | I'm digging up the Url |
| DELL-ToddM | http://www.dell.com/downloads/global/power/1q04-jav.pdf |
| DELL-ToddM | Lots has changed since then, but testing Vmotion is the common element |
| DELL-ToddM | I've been out looking on the web and on blogs for others running Sql Server on Vmware |
| DELL-ToddM | I |
| DELL-ToddM | I've found a few blogs and whitepapers from others |
| DELL-ToddM | I'll be posting these to the Sql and Vmware wiki page |
| DELL-ToddM | I've also done some work with a combined Emc, Vmware, Dell team on Sql Server |
| DELL-ToddM | The results of that work should be coming out soon |
| DELL-ToddM | Anybody have any questions? |
| DELL-ToddM | Darren - I saw you typing a few times, but nothing every showed up - Was there something you wanted to comment on? |
| DELL-ScottH | do you have any thoughts on how this might work with iscsi for backend disks ? |
| Dell-Darren | Well I was going to comment on some performance characteristics of Sql Server and how they apply in Vmware because that's one area where I haven't run any tests. You ended up covering most of what I had questions about. |
| Dell-Darren | In terms of running this with iscsi, the scaling should be similar. The one nice thing about fibre channel is you don't have to worry about the switches so much. |
| Dell-Darren | With iscsi you can hit potential bottlenecks at the switch infrastructure level. |
| DELL-ToddM | We have our Equallogic arrays on order - and I want to run some similar tests when they get in |
| Dell-Darren | I'd be very curious to the outcome of those tests. Unfortunately I do not have any servers here equivalent to the ones you used. |
| DELL-ToddM | I would also like to link to the new Dell Inside It blog (with Marc Farley as lead blogger) - http://direct2dell.com/insideit/default.aspx |
| DELL-ToddM | This is the first week - and it has had some pretty good posts |
| DELL-ToddM | Well, we have gone over the time by just a few minutes |
| DELL-ToddM | Next week's chat will be hosted by Scott Hanson and the topic will be.... |
| DELL-ScottH | Systems Management open discussion -- similar to a radio talk show with open callers ... should be fun ! |
| DELL-ScottH | unless we get prank callers :-) |
| Dell-MarcFarley left. | |
| DELL-ScottH | thanks for hosting Todd |
| Dell-Darren | Thanks Todd. |
| Dell-Darren left. | |
| DELL-ToddM | Thanks to everybody for coming |
| slbrown | Thanks for sharing your findings. |
| slbrown left. | |
| DELL-ToddM left. | |
| DELL-ScottH left. |