By Jeff Kabel, HP WW Technical Marketing Engineer, HP Networking
My mission: to prove that the HP Intelligent Resilient Framework (IRF) makes VMware more efficiently than a traditional design. The premise is simple: take a bunch of Virtual Machines (VMs) over here, use vMotion to move them over there, across an IRF-enabled network compared to a non IRF-enabled network.
Working with folks from Network Test, Inc., we created more than 100 VMs, setup our network topology, and started moving them. Easy, right? Not so much… Apparently vMotion is pretty smart and as VMs are running, memory that is not being used gets optimized. And when you move a VM that has been running for a while from one location to another—it does diddly to affect the network! Yikes! Punt. Try again.
We needed a big, fat, ugly VM to beat up the network and push bandwidth over 10Gbps. So we turned to the partner we knew would create a huge VM…Microsoft! Unfortunately after only 8 Windows 2008 VMs, our test bed wasn’t big enough to build enough of these to really tax the network. Suddenly my “easy” test was not looking so easy!
HP Proof of Concept Center to the rescue
The HP Proof of Concept (PoC) center in Cupertino, CA answered our call for help. The team there had all the same gear as our test bed. And servers! Oh boy did they have servers! So we packed up our stuff, kissed our loved ones goodbye and hit the road for another week of crazy testing. Once there, they gave us:
- 2 fully loaded HP C7000 Series blade enclosures
- 16 BL460 blades, each with 128GB of RAM
- More than 20 TB disk storage
- 2 HP 12500 10Gig core switches
- 2 HP 5820 10Gig ToR switches
- And a fully stocked coffee machine
Follow our testing path—and see what we learned
The first day was building out the VMs and VCenter Server. For the record, a big, fat, bloated VM takes MUCH longer to clone than a tiny little Linux VM. Coffee time! Then, after many hours of cloning, the VMs were built and the test bed ready for testing.
- Test #1 – Spanning Tree setup: One Active 10 Gig Link. We started the transfer of the now 64 VMs and were able to saturate the 10Gig pipe up to around 6Gig! Darn It!
- Test #2 – Spanning Tree setup: One Active 10 Gig Link We started the transfer of now 96 VMs and we see the traffic spike to 8Gig and then tank! What the. . .? Okay. Regroup. This is NOT as easy as it seems! After some reading (and a LOT more coffee), we realized that vCenter really DOES have a lot to do with the transfer speeds of vMotion!
To make a long story short. . .
- Test #93 – Spanning Tree setup: One active 10Gig link. We start the transfer of 96 simultaneous VMs, split across our now FOUR vCenter Servers! The pipe goes right up to 9.76Gbps and pegs for 15+ seconds! Woo hoo! Yes! High Five! After three day, we proved. . . STP sucks! Now on to IRF.
- Test #94 – IRF Setup: Two Active 10Gig links. We start the transfer of 96 simultaneous VMs! The pipe goes right up past 10Gig and then drops right back down to around 9.8Gbps! Come on!
To make a longer story even shorter…
- Test #127 – IRF setup: Two Active 10Gig links. Blurriness in both eyes and loss of feeling on my left side due to the 12 café mochas I just pounded at 2:00 am! We start the transfer of 128 simultaneous VMs. We now have an HP DL595 “Beast” of a Server with dual 10Gig NICs and 256Gig of RAM running as our vCenter Server. Bandwidth blows right past 10Gig. Yeah! Then levels to around 16Gbps. Argh! I tried creating more VMs but ran out of room on my storage (and could hear a small Scottish voice coming from my kidneys yelling “She can’t take much more captain!!” as I considered having one more cup of coffee!) That’s going to have to be it!
So what have we proven about the Intelligent Resilient Framework?
OK, IRF is better than Spanning Tree. Great. But did we prove that it was twice as better? No! Is that because it isn’t? Heck no! I think everyone would agree that having two links active is better than one.
So you are thinking: This is flawed test, or an “almost meaningless report”? I don’t think so. We’ve proven that IRF provides more bandwidth. If we had more time we could have built more VMs and added more servers.
We also demonstrated how—at no additional cost—IRF can give you a topology that better uses your available bandwidth. As added bonus, IRF has as fast, if not faster, convergence time than more complex, expensive technologies like SONET and MPLS TE!
So take it how you will, but know that HP’s IRF provides more for the same, or in most cases much LESS, than others out there!
>> Read my related blog post: Intelligent Resilient Framework rocks: boosting vMotion performance across the network.
>> Check out the Intelligent Resilient Framework video.