Wednesday, January 20, 2010

Intel, Microsoft top 1,000,000 IOPS in iSCSI tests

January 21, 2010 – In conjunction with Microsoft, Intel recently announced that it surpassed one million I/Os per second (IOPS) in an iSCSI performance test. See the Webcast link near the bottom of this post if you want detailed information on the tests and configurations, but basically the test configuration included a standard server with one Intel 10 Gigabit Ethernet (10GbE) server adapter (model X520-2), a quad-core Xeon 5500-series CPU, Windows Server 2008 R2, Microsoft’s iSCSI initiator software, and Hyper-V.

Other test elements included Cisco’s 10GbE Nexus 5020 switch connected to ten iSCSI soft targets.

Specifically, Intel and Microsoft clocked 1,030,000 IOPS (with 512-byte blocks), and more than 2,250MBps with large block sizes (16KB to 256KB) using the Iometer benchmark, which is waaaay beyond any iSCSI performance specs that I’ve seen.

It’s stuff like this that might bring back the old iSCSI vs. FC/FCoE controversy.

Of course, one million IOPS is overkill for most of today’s applications, so I queried Microsoft and Intel representatives about what these tests results really prove.

“These tests negate skepticism about iSCSI performance and its ability to be deployed in an enterprise environment,” says Sunil Ahluwalia, Intel’s product line manager, data center products. “They also demonstrate the head room you have with 10GbE and iSCSI.”

“From a virtualization perspective, heavily transaction-intensive applications such as databases tend to be the last applications to be virtualized, and these tests are a proof point for being able to use iSCSI in virtual environments with those types of applications,” says Dai Vu, Microsoft’s director of virtualization products and solutions marketing.

It was also a chance for Microsoft to show off some of the iSCSI and storage enhancements in the R2 version of Windows Server 2008, including iSCSI multi-core and NUMA I/O, DPC redirection, dynamic load balancing, storage I/O monitoring, CRC digest offload, and support for 32 paths at boot time.

For details on the iSCSI tests, see

the archived Microsoft-Intel Webcast

Intel’s blog

And in other one-million-IOPS news, Emulex recently announced that its FCoE-based OneConnect converged network adapters (CNAs) hit 919,000 IOPS in tests conducted by IT Brand Pulse. You can read the press release here.


McDizzle said...

This article seems to glaze over the fact that we aren't talking about legacy 10 gigabit ethernet. We are talking about Data center Bridging. The high IO via iSCSI is a result of the DCB and Priority Flow Control. This result couldn't be achieved on legacy 10 Gigabit Ethernet technology.

samecoin said...

The link provided to Microsoft's WebCast where there are supposed to be the nitty-gritty details on the testing opens a page with nothing of any use.

Any time these types of claims are made, it's vital to understand the exact configuration of every single component.
For example, did this test include Lossless Ethernet for the entire data path? If not, were any packets dropped during the testing?