AMD beats drum for Shanghai surprise
Anyone seen Big Blue?
Where are the tier one players?
The interesting bit about the system builder event was who showed up and who didn't. Penguin Computing, Colfax, Appro, Verari Systems, and Rackable Systems - all of who sell high performance clusters or dense clusters that are used for Web infrastructure and so-called cloud computing applications - got a chance to swear their fealty to AMD and explain how the Shanghai Opterons were going to give them an edge in the market.
There is no doubt a lot of good engineering at these firms, but none of the tier one players in the x64 server racket - that's Hewlett-Packard, Dell, IBM, and Sun Microsystems in that order - were at the event. That doesn't mean these companies are not going to sell Shanghai systems. All told, AMD says that 85 server motherboards and bare-bones servers have been validated so far to support the Shanghai chips.
The good news for AMD is that server buyers are "risk averse" right now, according to Banda, and that means selling upgrades to Shanghai processors or new systems that use Shanghai chips instead of dual-core Rev F or quad-core Barcelona chips should be relatively easy. With the future Nehalem processors, Intel is facing what could be a more difficult transition because the Nehalem chips and their QuickPath Interconnect is a radically different design from the front side bus architecture used for Xeon chips. Intel could have a smooth transition. It has happened before during technology transitions. Just look at the relatively painless move Apple made for its Mac PCs and Xserver servers when it moved to Intel chips.
While AMD rolled out these five server partners to talk about the ease of transition from Barcelona to Shanghai with simple BIOS upgrades, the fact is that the jump this time around was a lot easier than the jump from Rev F dual-core Opterons to quad-core Barcelonas. Almost by definition, it would have to be easier, and with Barcelona late (by six months if you want to be generous, more if you don't) and with a bug to boot, it would be hard to do worse with Shanghai than the company did with Barcelona. Nonetheless, it is important that the Shanghai chip doesn't yet appear to have any bugs, and that AMD was able to pull its production in from the first quarter of 2009.
Today, AMD also pulled a demo stunt that it is clearly hoping will be replicated in real data centers the world over. But based on market share figures and Intel's fairly aggressive Xeon server processor roadmap probably will not be.
Specifically, AMD teamed up with commercial Linux distributor Red Hat to show a live migration of a video streaming application running on a two-socket Xeon E5420-based server running Enterprise Linux 5 to another server based on the Shanghai Opterons also running RHEL 5 without any interruption to the video being streamed to a desktop. (You can see that video here on YouTube).
Red Hat did not say which virtualization hypervisor it used in the demonstration, but both the integrated Xen hypervisor and KVM support live migration from one box to another one. KVM supports the migration between AMD x64 processors and Intel x64 processors - as was shown in the demo - but according to Red Hat's documentation from 2007 provided thanks to the CentOS clone of RHEL, live migration using the embedded Xen hypervisor has to occur on two machines with the same processor architecture and virtualization extensions in the hardware (if the processors have any at all).
Red Hat's own and more recent virtualization guide, from January 2008, doesn't get into this issue with any specifics. This would seem to suggest that the Intel-to-AMD demo was done using KVM, but that depends on what the meaning of the word "architecture" is in this context. AMD was not able to clarify which hypervisor it used in the test before press time, but this mixed CPU live migration issue is something I will look into further. ®