From the ‘I can virtualize faster than you‘ files:
Red Hat has been partnering with chip vendors AMD and Intel for a long time. Every so often though Red Hat will announce something specific with one vendor (or the other) — that’s the case with the announcement today on a speed/functionality announcement with AMD’s processors running on HP hardware with Red Hat Enterprise Linux 5.
Red Hat is announcing that users can, “…achieve significant performance gains by
coupling new high-performance device drivers with the features provided
by Quad-Core AMD Opteron processors, available with HP ProLiant DL585 G5 servers.”
The improvements are non-trivial. By taking advantage of AMD’s silicon based Rapid Virtualization Indexing, the promise is that users will reduce the overall number of cycles required to enable virtualization.
The problem with virtualization has always been that it requires a certain degree of processor utilization which tends to impact performance such that often times virtualized application simply cannot perform at the same level as their non-virtualized counterparts.
The actual metrics reported by Red Hat show that in an OLTP (online transaction processing) environment test, with a 16-CPU system, there are considerable gains to be had. With a fully virtualized system running the Rapid Virtualization Indexing feature Red Hat reported a 21-fold performance gain over regular (non-Rapid) virtualization metrics.
“Red Hat and AMD
have worked very closely with the open source community to ensure that
full support for Rapid Virtualization Indexing is available with the
first Quad-Core AMD Opteron
processor-based systems to be offered by a leading hardware OEM,” said
Earl Stahl, vice president, Software Development at AMD
in a statement. “We’ve been able to ensure that customers can reap the
benefits of this new virtualization technology right away.”
The problem though, in my simple layperson’s opinion is that even though the AMD/HP/Red Hat metrics are significantly improved, they are still not at the same level as non-virtualized (that is native) environments. Red Hat reports that the AMD/HP/Red Hat solution, “…reached 77 percent of the
performance of a non-virtualized environment on one of the industry’s
most difficult database OLTP workloads.”
So in my opinion, while this is certainly a good piece of forward momentum news, for data centers with heavy OLTP workloads, the case for virtualization will still remain a utilization versus performance issue.
I would suspect given this massive leap forward, that others (be it IBM and/or Intel) will soon enough come out with their own tests that will continue to push Linux virtualization closer and closer to fully native performance levels.