« SOA at Work | Main | My First Web Server »
Virtualization Testbed
Suppose you had a couple of quad processor boxes with 16Gb of RAM each attached to a 1Tb SAN and enough VMWare ESX licenses to do whatever you wanted on the platform. What would you do? We’re wrapping up some initial VMWare performance studies on the boxes and I’m looking for ideas about what research projects to do with them next. I have a few ideas, but I’m curious if you have any.
Posted by windley on January 3, 2006 3:14 PM
Comment from Jared at January 3, 2006 4:11 PM
What about dumping the VMWare and trying the performance studies using XEN?
Comment from Fernando Medina, Jr. at January 3, 2006 4:46 PM
Why not try concurrent high IO sql queries on same physical boxes to see how vmware behaves sharing SAN connection.
Comment from Lee at January 3, 2006 6:43 PM
Might be interesting to play around with this: http://www.oracle.com/technology/tech/linux/vmware/index.html
Comment from Jim at January 3, 2006 8:38 PM
I'd like to see benchmarks between VMFS and RAW LUNs. At some point you break a size barrier where one makes better sense than the other. What is that magic number? Further more, benchmarks that define with specific numbers what qualifies as a "high I/O VM". VMware talks to LUN design with this vauge term but they never define what "high" is. Let's help them loose that ambiguity shall we?
Jim
Comment from mike at January 3, 2006 9:53 PM
I'd like to see multiple MS SQL servers, possibly clustered with high transaction loads to see how they perform.
A second thing would be to test the maximum number of vCPU VMs you can get with a specific load and the max number of 2 vCPU VMs. Vmware has done some of this for whitepapers but it would be nice to get a third part test as well
Comment from Fazal Majid at January 3, 2006 10:33 PM
It would be interesting to see if there are synchronization effects on the VMs similar to TCP global synchronization in TCP/IP networks, which reduces the efficiency of bandwidth utilization because all TCP connections get synchronized, ramp up, hit the wall and slow-start in sync, causing oscillations in bandwidth utilization:
http://en.wikipedia.org/wiki/TCP_global_synchronization
Comment from Tom Gordon at January 4, 2006 3:11 AM
I'd build a virtual GSP and see what you could do. Run up as many instances of Quake 3 or Counterstrike on the machines and see what happens when they get hammered with real-time traffic requests.
Alternatively host some MMO servers.
or pretty much anything to do with online real-time gaming :)
Comment from Eduardo at January 4, 2006 5:46 AM
I'd like to see the real performance improve between a single vCPU VM and a multiple vCPU VM.
Interesting is a comparation about XEN... XEN site has a benchmark, but they used vmware workstation 3.x
A couple Citrix Metaframe servers its curious to see in a stress test.
Eduardo Farias
Comment from Tarry Singh at January 4, 2006 7:01 AM
Phil,
I'd want you to try Oracle RAC on those boxes.
My ADvice:
Wait for ESX 3 and then go for 4-way Virtual CPU's 2 per box. 8G RAM per nodetwo Oracle Instances per box. Have two boxes. and have a highly available 4 node on two boxes.
and then performance test them, I'm willing to do it for you. I write (currently on oracle RAC) articles for DBAsupport.com and am on my way to test the configurations on old quad processor servers. I'd be nice if I could do it on some new boxes. This could even make it to the TPC!
Let me know if you want me to do it for you.
Tarry Singh
Comment from Chui Tey at January 4, 2006 8:31 PM
With Salesforce.com outages in-mind, investigate what kind of base platform services are required for hot backup of SaS.
For instance, after trialing salesforce.com, make it easy for a company to migrate the application back in-house, either through classic replication for data, and through JNLP-style redundancy for applications.
Also, provide failure scenarios where internal server fails, and failover is provided by SaS on-demand.
We need sound theory on how data and services can be hosted and replicated. To this end, I see virtualization as a key component of hosted mobile application servers.
Comment from Andrew McCreath at January 5, 2006 5:49 AM
Why not start the first virtualised GRID computer, see what you get out of it.
I would imagine you could leverage more cpu processing cycles...
Comment from Nathan at January 6, 2006 4:23 PM
Another possibility would be testing the SMP/threading abilities of different os/kernels under a highly-threaded, processor intensive load.
Comment from Pat at January 6, 2006 8:38 PM
We've used similar setups for virtualized labs, where students need only have what's needed to connect to the VMWare server box and run the entire spectrum of lab capabilities (multi-O/S routing and service sharing) without having to install anything. We're testing to see if the VMWare Player will afford the same type of capability...
Comment from Raymond C. Parks at January 9, 2006 12:00 PM
We are building exactly such a system for the purposes of virtualizing control systems. Testing of security of control systems is problematic. One dare not test the real control systems for fear of causing the very consequences one is trying to prevent. Yet laboratory systems are too small to match the complexity of the real world systems. We are compromising by using VMWare and powerful servers as a middle stage in our virtualization of control systems. This middle stage fits between actual hardware/software and completely simulated hardware and software.