RSS
Email
Print
ShareOne Million IOPS: Moving The Data Center Bottleneck
Posted by Frank Berry on March 9, 2010
The first system administrator to ever upgrade a server undoubtedly experienced a mismatch in performance between the installed server and the new component. Since then, millions of data center managers have experienced first-hand the performance bottleneck constantly moving from one component in their data center to another. Have those years of experience stopped smart people in our industry from challenging whether it makes sense to deploy new generations of faster technology, because the new technology has performance the rest of the system can't use? The answer is no. The latest generation of high-performance converged network adapters is a great case-in-point.
Recent press releases, test reports and webcasts from Broadcom, IT Brand Pulse, Intel and Microsoft have included references to performance of one million inputs/outputs per second (IOPS). The reflex reaction from many in the industry was, "this level of performance is meaningless." The reason they believe one million IOP performance is meaningless is because is most servers, storage and applications can't use that much I/O capacity. They're absolutely correct about the capacity, but they're wrong about the additional performance being meaningless. All things being equal, data center managers prefer to buy the newest, fastest equipment in order to future-proof their environment.
If you're shopping for a car and the sales person offers you a choice of a Gremlin and a Porsche for the same price, are you going to argue the additional performance can't be used on the roads you drive, and choose the Gremlin? No, you're going to drive away in the Porsche and let the other poor S.O.B. explain why he bought the Gremlin. Similarly, in a recent survey of over 100 IT professionals, 77 percent said, all things being equal, they would buy Adapter A that did 1 million IOPS and was twice as fast as Adapter B. Furthermore, 36 percent of the respondents said they would even move off the standard adapter platform to get the additional performance.
In 2002 LSI announced the ability of their Ultra320 SCSI host bus adapters to support 100,000 IOPS. Their competitors rightfully pointed out that no server or storage could feed the beast with that much I/O. However, the product was successful because storage administrators wanted the fastest storage controller they could get. In 2010, 10Gb network adapter manufacturers need to achieve 1,000,000 IOPS performance because converged network administrators responsible for virtual server environments will expect it. And around 2015 when 100Gb Ethernet becomes mainstream, I expect many will challenge why anyone will ever need 10 million IOPS.
Comment by Mike on March 9, 2010 1:45 PM
Frank,
this somewhat misses the point. Popular applications typically operate with 4KB or 8KB block sizes, so they will saturate a 10G link with about 240K IOPS (using the smaller 4KB block size). As long as the adapter has headroom above this you are fine. Getting to 1M IOPS with some smaller block size is not useful.
As far as your Porsche versus Gremlin analogy, all things are never equal - there are differences in other performance metrics (eg, latency), power consumption, virtualization capabilities, scalability, integration/footprint, etc.
Reply to this comment
Comment by Frank Berry on March 9, 2010 6:15 PM
Mike - I agree with everything you say about the block sizes not being real-life, saturation long before 1M IOPS, etc. But I respectfully suggest that it is you that is missing the point that IT people regularly buy the faster widget, even if they can't use the perforrmance (today). This IT buying habit is every bit as real as the accurate points you make. Like my company tagline says, "IT perception is reality." -Frank
Reply to this comment
Comment by Anonymous on March 16, 2010 11:57 AM
Frank,
You are right on - the history of technology, especially IT technology, has always been about "faster/better/cheaper".
It is a Moore's law imperative.
Yes, there are those ostriches, with their heads in the sand arguing that "nobody needs/asks for this much performance". But look at Ethernet (1 Mb/sec to 100Gb/sec), modems (300bps to 56Kbps), SATA (now up to 6Gb/sec and giving FC a run for its money).....
Industries reach Clayton Christensen's "inflection points" - and those industry participants suffering hubris, and refusing to acknowledge such paradigm shifts, are eventually consigned to the dustheap of technology history !
Reply to this comment