« Moving right along.... | Main | The Quest continues.... »

November 14, 2007


Feed You can follow this conversation by subscribing to the comment feed for this post.

open systems storage guy

Concerning virtualization, Xen (and now Oracle) have to catch up to a 3 year lead VMWare has on partner and interoperability development. VMWare's biggest competitive advantage (and the reason I don't think hypervisors will become a commodity any time soon) is that for 3 years while they were the only game in town and were sitting primarily in test centers, they were refining their ability to work with the majority of commercial server software. Sure, the other guys can do the easy stuff, but as soon as someone wants to put something besides a file or email server into Xen, they will run up against the lack of proper development for interoperability.

Concerning benchmarking, I think that an open sourced load generator would be ideal- if anyone finds a characteristic of the workload that's unfairly benefiting one type of device, they can fix it or fork it, depending on whether it would be useful to show both.

I don't know if this is feasible, but couldn't we just record IOs from a production environment and play them back to the devices being benchmarked?

Subramanian Kartik

On the Xen/Viridian coming of age - in complete agreement. Seeing where VMWare is today, its a clear 2-3 year lead. No doubt the competition will make inroads, especially, IMHO, Microsoft, but not anytime soon.

On benchmarks: The critical piece is that it be representative of real workloads. Which is why I am in favor of what I espoused in Postulate #1. A single measurement will never reveal the true nature - strenghts and weaknesses of a given platform configuration. So multiple different workloads should be run against the same configuration - that way, if the configuration has been tweaked or shows a benefit for one kind of workload, that will wash out in the composite picture. Say small block/large block, cache-friendly/cache-hostile, sequential/random. Hard to simultaneously optimize for all of these in the same configuration.

Pat Artis has tools that can characterize real workloads on the mainframe by capturing and modeling real life IO streams - haven't seen any similar tool for open systems. Has anyone?

Cheers, K.

The comments to this entry are closed.