Thursday, November 4, 2010

Intel’s card play in unified networking (10GbE+iSCSI+FCoE)

November 3, 2010 – InfoStor has been covering converged, or unified, networks and the Fibre Channel over Ethernet (FcoE) protocol for years, but most of our product-oriented coverage has been centered on adapters from vendors such as Emulex and QLogic. Those vendors are clearly out in front in this space, both from a time-to-market and market share perspective, but there are other vendors with host adapters for converged networks, and we’ve been remiss in not covering a little player called Intel.
First, a note on terminology: Vendors such as QLogic and Emulex use the term converged network adapter (CNA) to describe cards that support protocols such as 10GbE, Data Center Bridging (DCB), iSCSI and FCoE. Other vendors, such as Broadcom, are expected to use the term Converged Network Interface Card, or C-NIC.

Intel doesn’t use either of those terms, referring to the technology in general as “unified networking” or just referring to its Ethernet X520 Server Adapter card.

Intel’s approach to unified network cards is architecturally different from the CNA approach taken by vendors such as Emulex and QLogic. With CNAs, protocols (and protocol offload functionality) are implemented on the CNA, which architecturally is similar to a host bus adapter (HBA).

In contrast, Intel uses native initiators for FCoE and iSCSI that are implemented in the host operating system (Windows and Linux) kernel. Thus Intel’s use of the term “open FCoE.”

“We rely on native protocols in the OS kernel, so it’s free and easy to implement and use,” says Sunil Ahluwalia, senior manager of product marketing in Intel’s LAN Access Division. “CNAs are special-purpose adapters that are more complex and costly.”

Intel’s Ethernet X520 Server Adapter has an MSRP of $799 (and often sells for considerably less), which is about half the price of CNAs.

Rather than implementing hardware-based full offload of the storage-over-Ethernet protocols, Intel offloads only the data path part of the protocol.

“We have what we call ‘intelligent offload,’” says Ahluwalia. “We don’t offload the full protocol because we don’t think that’s required, and full offload adds complexity and cost to the cards because you need processing engines, memory, etc. on the card, which also adds power requirements.”

Performance

CNA vendors have historically claimed that the native initiator approach (a) lacked full support for protocols such as FCoE (which is no longer the case) and (b) had relatively poor performance.

Intel begs to differ. However, it’s important to note that all vendors in this space can point to internal or third-party testing that shows superior performance for their products. As always, benchmark results should be taken with a grain of salt.

“Intel and Microsoft demonstrated one million IOPS with native iSCSI earlier this year, but the question was: Who needs that level of performance in real-life applications?,” says Ahluwalia, “so we performed some tests [with Demartek] running applications such as Exchange and SQL and we found little difference in performance between a CNA and an initiator approach.”

For full info on the testing procedures and results, see Demartek’s “Intel 10GbE Adapter Performance Evaluation for FCoE and iSCSI.” (Demartek’s site has a lot of other interesting FCoE-related content).

“We found that the performance of the Intel X520 adapter was comparable to competitive 10Gb FCoE adapters for a broad spectrum of tests,” said Dennis Martin, Demartek’s president. “Because the adapter performance was reasonably close in most of these tests, IT professionals need to consider the cost of these adapters, especially in environments where many adapters are required.”

Intel’s Ahluwalia also claims that the native initiator approach does not consume a lot of CPU cycles (another criticism of this approach).

To take one example from the Demartek tests: In a Microsoft Exchange Jetstress benchmark with 5,000 mailboxes, Intel’s card/initiator consumed only about 2% CPU utilization.

But you really have to take a close look at the Demartek tests, as well as testing of CNAs conducted by other vendors and third parties, to get to the bottom of the performance claims in the unified, or converged, network adapter market.

Related article: QLogic announces 10GbE NICs, CNAs

2 comments:

Unknown said...

Dave,
The question I have for Intel isn't so much performance or cost, but reliability and interoperability. Until the OS vendors and storage vendors have qualified/certified the solution, customers will not use it for FCoE environments. The Fibre Channel driver stack is "sticky" with many enterprise customers who want to have their entire application stack tested before deploying. Intel needs to prove that they can support large-scale enterprise deployments. That being said, Intel is well known for general NIC use and iSCSI and will be a tough competitor in 10GbE. Also, Intel today only has Windows and Linux, a big gap is VMware (Emulex and QLogic CNAs are already supported in VMware and other UNIX OSes).
Thanks,
Stuart Miniman
http://wikibon.org
Twitter: @stu

Dave Simpson said...

Thanks for the input, Stuart. I think the OS (Windows, Linux) vendors have in fact qualified the "open FCoE" stack. However, Intel has so far only qualified the stack with storage platforms from NetApp, with EMC qualification due within the next two months.
Good point re Intel only supporting Windows and Linux, and not VMware. -Dave