ORIGINALLY POSTED 11th September 2007

10,475 views on developerworks

I spotted Hu has made a long post about the new mid-range USP, and various articles out there are discussing it. However, it seems my replies to both Hu and Claus fell on deaf ears – yet again they are quoting those incredulous attachment figures.

So yet again lets look at the numbers Claus and Hu are banding around : (I already discussed the numbers over at ‘zilla’s place)

The USP VM can also virtualize up to 96 Pbytes of heterogeneous external storage compared to the USP V’s 247 Pbytes, he said.

Being kind, lets chose the largest, slowest disks : 750GB SATA which can do about 80 ops each. 96PB is 128,000 of them – lets assume the RAID is already taken care of in the attachement figure. So 128,000 x 80 = 10.24 MILLION IO/s

The only figure HDS are quoting is the 1/3 performance of USP-V, or 1.2 Million “super read cache hits”. By super-read cache hit, what I mean is a frame that never leaves the FC port buffers. The HDS products have this special ‘unrealistic of real life’ mode, where they can return data if its still in the frame buffers – think of it as cache on an HBA. So thats not even entering the system processing complex. (FYI: Any SVC cache hit numbers you see are actual data coming from the cache) Anyway, 1.2 Million is a long way short of 10 Million. So while the “addressability” may well be able to cope with 96PB of external storage, just how many of them could you use?!

If we go for 500GB 15K RPM FC disks, then we are talking about over 38 MILLION IO/s capability…

Anyway I’m beginning to feel like a stuck record, and hopefully I’ve made my point. Oh and one more thing, while SVC’s hardware base is indeed a server, why not? Its got seriously good bandwidth into the memory / CPU complex, the leading edge in PCIe bandwidth and comes pre-tested, pre-developed and a lot cheaper than building custom hardware every 2-3 years that uses last years technology because its all that was available when the hardware was being designed.

All that said, I like the idea of a virtualizing controller (personally) but hey, buy an SVC, stick some relatively cheap SAS or FC RAID behind it and et-voila, you have one for A LOT less of the green folding stuff. It will have more performance capability, can be migrated to next years model without any downtime and can be scaled-out as and when you need it. Just a thought…