A sneak peek at vSphere 4 I/O goodness (and world records)

So, about a year ago (May 2008), the VMware Performance engineering team and EMC collaborated on an effort to see just how much IO you could push through an ESX 3.5 host.  The results of that effort were posted in detail here, and it started a lot of dialog about what applications were candidates for virtualization.

Some of you might have seen this IOmeter “speedometer” that represented the maximum IOps (100,000) we were able to push through a single ESX 3.5 host.

For perspective – this took 3 CX3-80s, 500 15K RPM spindles – and is the IO workload equivalent of roughly 200K Exchange 2007 mailboxes, or 85 moderate SQL Server databases.

At that time, we were saying together that almost all x86 applications were virtualization candidates, and over the last year – more and more customers have been embracing this idea.

At that time, a good question that represented a barometer of “virtualization sophistication” at a customer was “are you using DRS, and if so,are you using it in fully automated mode?”  If they answered positively – it meant the customer was embracing this idea of “infrastructure automation” and was building trust in the VMware datacenter layer.   Today – the analagous question is “do you have a virtualization first policy, and how far along are you with virtualizing Exchange, Sharepoint, your biggest SQL Server and other mission-critical applications?”

This is a more important idea than it might be on it’s surface.   The first step of getting to a federated private cloud is to build an internal cloud.   Building an internal cloud where infrastructure is virtualized infrastructure and IT is starting to trust this new datacenter layer with the crown jewels.   Put differently – the first step to the private cloud is the 100% virtualized datacenter – all apps, all workloads, all use cases.

The reason we do these crazy workloads (including some of the others you saw today) not because these are common – but precisely because these workloads are uncommon – to banish the idea that “well we can’t virtualize workload ______” – and start with the question instead as “why can’t we virtualize it (and ‘big performance’ isn’t a reason)?”

So it was inevitable that with vSphere 4 approaching, we would repeat the exercise.   All the detail will come around EMC World (register to see these details and MUCH MUCH MORE!) and will be posted on the VMware VROOM blog.   The engineering teams are professionals, and as much as I tried to cajole them, they don’t want to publish them until every T is crossed, every I dotted, but I’m so pumped about the results, I have to share some of them.

• we CRUSHED the previous “single ESX host” record – we’re at 325K and we still haven’t hit the limit – we’re still climbing!  WOW.
• this test was conducted in a variety of configurations and I/O workloads – more details to come, but no, it didn’t require the use of VMDirectPath.
• The ESX and guest latency?  Too early to publish, but WOW.
• The new pvSCSI vSCSI adapter is showing that it can deliver substantially more with a given CPU utilization (and of course same IOps at substantially lower CPU utilization).   WOW.
• We demonstrated that ONE CX4 could do the work of 3 CX3-80s in the previous year. WOW.
• While the story is absolutely a vSphere 4 story, a cool storage-side equation that shows another aspect of being efficient - we demonstrated that 30 of the EMC EFDs in the CX4s could do the equivalent work of 2000 15K Fibre Channel drives.   That’s 85% cheaper acquisition cost, and 99.5% less space and power,  WOW.

Same rules as last time – this is an extreme demonstration.    COME TO EMC WORLD TO SEE ALL THE DETAILS! 

But – when VMware says vSphere 4.0 is ready for any application, any workload, consider this datapoint.  It’s an exciting new day.

Comments

[duncan]

325K wow, that's insane. Would be nice to see what the difference is between the pvSCSI and the "old" scsi drivers.

[Daniel]

Great right up chad, its definately evident that work and orchestration you and your team are doing between EMC and VMware is actually working wonders on workload performance improvements...

I can just see this getting a whole load better with introduction of external infrastructure such as Vmax, Cisco UCS/Nexus etc all working in tandem with vsphere..

[Charlie Dellacona]

325K IOPS! Wow indeed. Congratulations.

It would really drive home your point if you published the exact same test without ESX on the server. That would not only settle the scale argument but demonstrate the efficiency as well.

[Keith Norbie]

Super cool, nice use of all the spare time! Does the network matter? iSCSI, FCOE, FC, NFS? Test through NS not just CX4. It would be cool to also see F.A.S.T. working to distribute workload and capacity demands to really attack the legacy datacenter array footprints.

(disappointed, always serious - insert "bork" somewhere... spread your Canadian wings)

[Chad Sakac]

@ Duncan - will send it to you on the down-low (as a VMware employee) - it will be posted shortly for all to see (after EMC world - the teams want to finish up all the testing).

@ Keith - just for you... BORK :-)

[Shawn E]

Would it be possible to perform the same test with Hyper-V?
Even though there was a (controversial) article written where it came out ahead of ESX in a very limited, low-end scenario, I doubt it would scale the way the new ESX does.

[Chad Sakac]

@Shawn E - I'll find out from my Hyper-V peers what they have planned. I'm sure that Microsoft, seeing all this will want to work with us to do something similar, but that's not in my part of the org...