IBM preps 32-way Intel server
Truly big iron
The 32-way server will initially use the 32-bit Gallatin Xeon MP chips recently announced by Intel, and will, according to IBM sources, eventually support Intel's 64-bit Itanium processors - possibly the current McKinley Itanium 2 chips or their follow-on, the Madison chip, which is expected sometime in mid-2003. The code-name of the 32-way Intel box is as yet unknown, but sources say that IBM will not brand it under the xSeries 440 moniker.
The forthcoming xSeries machine will give Big Blue a high-end server which can compete against the current line of Wintel servers from Unisys and the future Pinnacles chipset from Hewlett Packard. The latter will scale from eight to 64 processors, and eventually to 128 processors, in a single system image. It is unclear if IBM plans to take the kicker to the Summit or EXA2 chipset to those scalability heights. But the company undoubtedly wants to build on its momentum in the high-end Wintel market, and with Windows .NET Server 2003 expected to be ready in April 2003, IBM wants to have a machine in place that can really take advantage of that operating system.
The 16-way "Man-o-war" xSeries 440 should start shipping early next year; it has been delayed mostly because of Microsoft Corp's continuing delays with the kicker to its Windows 2000 operating system, once code-named "Whistler" and now to be called Windows .NET Server 2003. Whistler is important because it will be the first real full-blown MS product to support 64-bit processing, large memories, and NUMA clustering - something that IBM and HP have wanted for their Wintel servers for years.
Intel's delays with its Itanium 2 and Xeon MP processors - both of which underperformed compared to expectations - has pushed out the 16-way xSeries 440 deliveries as well, which were expected in early summer. Without a decent chip and an operating system - Linux does not scale to 16 processors (excepting on specialized HPC boxes like those sold by SGI), and probably will not for years to come - it's been hard for IBM to get the xSeries 440 up to its full scalability.
While the xSeries 440 is based on a four-way cell board clustered with a high-speed switch and NUMA memory architecture to create a single system image, the future 32-way xSeries machine will be based on eight-way cell boards that will probably also be sold as standalone eight-way servers in the xSeries line.
The initial xSeries 440 machines were based on a variant of the Summit chipset that only supported the Foster Xeon MPs, which topped out at 1.6GHz. The version of the Summit chipset used in the 16-way xSeries 440 will support the Gallatin Xeon MPs and the McKinley Itaniums, which should give the xSeries 440 a significant scalability boost.
The Gallatin chips top out at 2GHz and have twice the L3 cache memory as the Fosters (either 1MB or 2MB, compared to the Foster's 512KB or 1MB). The Gallatins also have support for HyperThreading, which can boost performance for applications that have be tweaked to support it. It is unclear if the early versions of Whistler server and the versions of SQL Server 2000 that vendors are playing with support HyperThreading.
Based on raw clock speed alone, a machine using 2GHz Gallatins should run 25% faster than one using 1.6GHz Fosters; the larger L3 cache on the Gallatins could easily boost performance by 10% or so. HP has tested its F8 ProLiant DL-760 G2 machine using the 2GHz Gallatins using a beta version of Whistler server and has been able to hit 111,805 transactions per minute (TPM) with the machines on the TPC-C online transaction processing benchmark test.
An eight-way Pentium III Xeon using 900MHz processors will hit around 70,000 TPM, an eight-way using Foster MPs will hit around 90,000 TPM. The ratio of performance of the F8 server compared to a Foster server with eight processors suggests that HT is not turned on, since the Gallatin machine only yields about 38% more throughput compared to IBM's Foster xSeries 440. The differences could be explained by the differences between the IBM Summit chipset and the HP F8 chipset, of course. It will be hard to tell until IBM puts out its own xSeries 440 benchmarks using the Gallatin chips.
It's hard to tell where 16-way and 32-way Gallatin machines might top out, but a 16-way could be in the 175,000 to 180,000 TPM range if Whistler server in the Datacenter Edition can scale as well as Unix servers on OLTP jobs. Benchmark results from Unisys suggest that this kind of scalability is possible with Whistler and Xeon MP processors. A 32-way ES7000-230 "Orion" server using the 1.6GHz Foster chips hit 203,518 TPM on the TPC-C test recently running Whistler and SQL Server 2000. With 2GHz Gallatins, that 32-way would probably hit around 280,000 TPM, and it might even hit 300,000 TPM with tuning and HT turned on.
While vendors quibble over the differences in their 32-bit, eight-way machines, customers looking at Wintel and Lintel applications will be watching the Itanium 2 and future Madison chips very closely. An eight-way using 1GHz Itanium 2 processors should hit around 125,000 to 150,000 TPM, depending on the chipset. This is one of the reasons why IBM has Summit chipsets for both 32-bit Intel Xeon MP chips and Itanium chips. With a 1.5GHz Madison chip, an eight-way could hit 180,000 TPM to 225,000 TPM, and a 16-way machine might get close to 350,000 to 375,000 TPM. A 32-way Madison machine, if such a behemoth could be built and tuned for Windows .NET Server 2003, would be in the 500,000 to 600,000 TPM range (there's some wide error bars on those estimates).
This is truly big iron, and there's nothing that those in the Unix camp will be able to say against it - particularly if IBM and HP can deliver such machinery to the market. As has been the case for the past decade, this can only happen if Microsoft and Intel deliver on their processor performance and operating system scalability promises. Given that, it is understandable that the Unix vendors - including IBM and HP - are not exactly shaking in their boots.