Aidan Delaney’s weblog

I was at an IBM sales conference yesterday. It was very interesting and targeted migration from Solaris to Linux on OpenPower. I know I won’t be on the Moulsecoomb campus today so I’ll post this (I’m writing off-line) from the Falmer campus. Which basically means that instead of emailing interested parties (like I usually do) the whole world can Share My Wisdom(tm). Update: There’s no wireless on the Falmer campus. I had to wait untill this morning to post this from a coffee shop.

I’ve been a closet fan of the Power architecture for a while now. I’m not quite sure why because I’ve never had much of the hardware, just a few old PowerPC G3s that people throw out and I hoarde. I suppose it comes from seeing what a high-calibre RISC chip like the PA-RISC and Sparc chips can do that Intel still can’t (or don’t at commodity price points). It may also have something to do with reading the fscking manuals. I’ve read books 1 through 3 of the Power architecture. The platform appears clean, scalable and legacy free (when compared with x86). I suppose it’s now the only game in town after HP committed infantacide with PA-RISC and simply murdered Alpha. Itanium still hasn’t proven itself and is more expensive than OpenPower and finally Sparc, for all it’s open specification, seems to have no champion with a fab plant. Neither TI (for Sun) or Fujitsu/Siemens have a credible roadmap that I’ve seen. So I went in a fan, particularly of the operating system they were pushing

From the presentations and the demos it seems to be a serious beast. I was mightily impressed by what I was told could be done with it, such as really fine grained virtualisation that is not possible with Xen on Intel. At that, the fine grained virtualisation is (reputably) faster than Xen on Intel as it leverages the advantages of OpenPower architecture (btw it’s a hypervisor [which I think is IBM-speak for an exo-kernel] rather than para-virtualisation). Stuff like sharing fractions of a CPU, or my favourite: allowing one partition to monopolise the vector processing unit (I assume is an Altivec derivitive) whilst sharing the ALU and FPU with other partitions. Think about running Apache (no vector processing) on one image and a hard bioinformatics problem on the image “owning” the vector registers. In Intel speak it’s ensuring one virtualised image on Xen owns the MMX and SSE units and shares the multiple CPU ALUs with other virtualised partitions. It seems to have some pretty awesome power management aswel, such as being able to turn off CPUs that are not in use, thus saving electricity and air-con costs. But most importantly, and this I learned today, it improves the utilisation of the system.

The argument is, why have two servers doing nothing for nearly all their life (print servers do nothing once the office is closed) when you can consolidate them on a single machine as securely separate virtual images. Basically if you run a machine at 100% for 100% of it’s life you’re getting a better return on investment than if you had spare capacity. This all makes sense in hindsight, but I’d never heard that argument before. It’s also one of three reasons I’d caution against splashing out on OpenPower at the moment.

Reason one: Unless I can prove that I’m using the current hardware to practically 100% utilisation why should I upgrade? Well I can think of answers to this question such as, It’s cheaper to throw more CPUs at a problem than have me work on it for a week (after all, I am expensive . But realistically I really should have some proof as to my current system utilisation before I consider upgrading, how else can I know that my potential new system will solve my problems? Anyway, the polite, knowledegable IBM sales lady has only herself to blame for this reason. Everything I’ve just written is a regurgitation of what she said.

Reason two: GCC does not fully support the vector units. Why move to OpenPower when I get good SSE2 code generation on Intel? Of course this depends entirely on the type of job to be processed. This argument simply doesn’t hold in the case of a web server/print server, however it does hold for scientific data processing.

Reason three: The IBM hypervisor is not open source and will never be open source. This is a really big limitation. “Pragmatists” out there will probably disagree, however we’ve worked too hard to get propritery software from between Linux and the hardware to let IBM insert a new layer. Plus I don’t trust them If I were them I’d grow the Linux on OpenPower market and then lock Linux out from the really cool features of OpenPower “enticing” (read forcing) customers to migrate to the “better” proprietary UNIX i.e. AIX. If I’m that evil I’m sure they can be too.

So there you go. I’m going to have a look at measuring the level of utilisation on my servers and then if I find them under-utilised I’ll virtualise it with Xen…just for the hell of it The sales people were very nice and extremely knowledgable (you’d expect that though). They were also very candid, explaining where Linux on OpenPower may loose to Linux on Intel or Solaris on high end Sparc boxen. Plus it was really cool to finally meet Jono Bacon who was the speaker on open source.

After that post I bet you’re all feeling hungry. Here’s a nice recipie for muffins.

• 300g of flower
• 2 tsp baking powder
• pinch salt
• 2 large eggs
• 100ml milk
• 75g sugar (I use brown cane sugar)
• 75ml vegetable oil

Add the first 3 ingredents into a bowl. Combine second 4 into a bowl together. Throw both bowls in together, mix it up. Fire in some soft fruit (I use strawberries) and some white chocolate (Green&Black organic chocolate is nice). Mix a little more and put into cases in a 200 degree centigrade oven for about 10-15 minutes. Enjoy.

This entry was posted on Saturday, July 16th, 2005 at 1:58 pm and is filed under General. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.