I've just finished reading a couple of Intel papers on I/O AT from a networked storage point of view and wonder whether Intel is, to be blunt, not grasping the right end of the stick? Intel's Pat Gelsinger, senior vice-president of Intel's new Digital Enterprise Group, said this regarding I/O AT: " "We looked at TOEs [network interface cards with TCP/IP offload engines] and found they did no good in most cases. I/O AT increases data exchange by at least 30 percent for many different kinds of server traffic." Exit TOEs; enter I/O AT.
Just to remind us what I/O AT is about; with it the TCP/IP protocol stack is optimised for the (Intel) processor, copying of packet header and payload data is supported in an Intel chip set, with parallel processing and direct-memory access being a feature of an Intel network controller. Intel is reported as claiming: 'putting the TOE capabilities inside a chip instead of relying on a separate card increased performance in 90 percent of the applications it tested."
So TOES are not necessary. Or are they?
The Intel white paper
The first paper is 'Accelerating High -Speed Networking with Intel I/O Acceleration Technology'.
In it the Intel seems to be saying that TOEs are not needed because the actual CPU overhead involved with TCP/IP processing is small. Most of the time is spent fetching and copying packet header and payload data. But it also says that TOEs are effective with data warehousing and 'high-performance databases', also storage backup and retrieval and 'enterprise databases' - all of which use large data payloads with packets of 8KB or more.
Intel asserts TOEs are not needed for smaller network data payloads, web services, messaging traffic, communications within clusters, RFID and real time data with data payload sizes typically less than 1KB. Unless you are doing backup or enterprise database work you don't need TOEs; that's Intel's message.
But, correct me if I'm wrong, no one in the storage industry is suggesting using TOE NICs for RFID or messaging or, indeed on client PCs, as Intel clearly implies; Frank Spindler, VP of technology programs for Intel, has said that I/O AT can improve interaction between networked clients and servers by up to 30 percent.
No storage industry supplier is suggesting that TOEs should be used to connect clients to servers. The storage industry says TOEs are used to connect servers to networked storage and that means iSCSI SANs or NAS over Ethernet.
And, correct me if I'm missing the plot, but aren't virtually all networked storage data accesses about packets substantially larger than ones with less than a 1KB data payload?
Alacritech spokesperson Valdis Hellavik said: "Your thoughts about Intel are right on the money ... they're right on target."
So what is Intel playing at? Is it playing up an I/O acceleration technology as being generally applicable to storage access when it fact it doesn't apply to networked storage access at all?
The second paper is 'Intel I/O Acceleration Technology'. In it Intel talks about increasing storage I/O performance and says a key element of its I/O AT solution includes use of a 'storage processor' (the new Intel IOP333 storage processor) to retain this workload as a peripheral function'. The workload being data transfers to and from an application. No performance numbers are given. So we might assume that such a storage processor does not in fact increase server networked data access over Ethernet at all. It's only pertinent to direct-attached disk performance.
Direct-attach disk acceleration
Intel's web site states: ' Intel Application Accelerator has the ability to substantially increase the storage subsystem speed by increasing the data rate, from the disk to the system, using special software technologies. Consequently, all boot processes, except for the BIOS, are inherently accelerated.' So access to directly-attached disks can be accelerated. But this Application Accelerator does not apply to networked storage access and is separate from I/O AT.
TOE card review
Techworld reviewed a iSCSI TOE card from Adaptec. Here's what our review said about CPU utilisation: 'Adaptec scores heavily when it comes to processor utilisation. For the single disk and four disk striped array tests Iometer reported CPU utilisation as 18 per cent and 19 per cent respectively when using the Microsoft (software) initiator but with Adaptec in charge these figures dropped to only 3.1 and 3.5 per cent. We saw the biggest improvement when running the same tests on a Pentium III 866MHz system loaded with Windows 2000 Server which saw a 25 per cent reduction in CPU overheads."
The reduction in CPU burden was substantial. This, on the face of it, doesn't square with Intel's assertion that actual CPU processing is only a small proportion of TCP/IP packet processing.
Intel's own TOE
It might surprise you but Intel has its own TOE product. Its PRO/1000 T IP storage adapter competes with iSCSI TOEs from companies such as Adaptec and Alacritech. Blaine Kohl, marketing director for high-end server connections in Intel's LAN Access Division, is reported as saying the PRO/1000 T uses a full-TCP/IP offload engine design in which TCP/IP and iSCSI protocol processing is offloaded to Intel's IOP310 chipset (based on Intel's XScale architecture) and 82544EI Gigabit Ethernet controller. Intel claims CPU utilization of 3% to 5% or less from its own testing - that roughly concurs with our test of the Adaptec card.
Clearly the IOP310 functions as a TCP/IP offload engine (TOE). Has Intel now disowned its own TOE? Is its I/O AT a TOE amputation message?
I/O AT and iSCSI acceleration
It appears likely that Intel's I/O AT is simply not applicable to the iSCSI storage market. Contrary to Intel's general message TOEs are still most definitely needed to accelerate I/O over networked storage links. I/O AT may well help networked clients talk more quickly with application servers but that has nothing to do with networked storage access. Intel has not been able to respond to questions on this matter.
The iSCSI storage suppliers can use their TOEs to give Intel I/O AT a boot into touch, until and if Intel comes up with something that truly replaces TOEs for networked storage.