We are building a new version of a JackRabbit for a customer. During some of our testing, we booted it running SuSE 10.2 diskless, and I ran a GAMESS benchmark. We have been using this for years to exercise machines, and get rough performance comparisons. On 4 CPU Opteron 275’s with 8 GB ram, it takes ~3 hours. So I ran it recently on an AMD 2350 2.0 GHz quad core and our new JackRabbit.

As noted a few times here, iozone as written, has numerous hard-wired quantities within it, some of which impede testing for large and fast systems. Here is a simple patch to fix one of the issues … diff -uNr iozone3_283/src/current/iozone.c iozone3_283.new/src/current/iozone.c --- iozone3_283/src/current/iozone.c 2007-02-19 12:12:18.000000000 -0500 +++ iozone3_283.new/src/current/iozone.c 2007-09-12 11:47:19.000000000 -0400 @@ -741,7 +741,7 @@ /* At 16 Meg switch to large records */ #define CROSSOVER (16*1024) /* Maximum buffer size*/ -#define MAXBUFFERSIZE (16*1024*1024) +#define MAXBUFFERSIZE (1024*1024*1024) #endif /* Maximum number of children.

We submitted a bid for a cluster. A large one, and we were being very aggressive on price. Very thin margins, spoke with our suppliers to make sure we got the best deal we could get. Come the bid open and … we are on the high side. Some of the bids are lower than our cost of materials. Ok, if everyone is bidding the same thing, how is this possible?

My friend Doug O’Flaherty has a new blog. I won’t mention Doug’s affliation, as this is not actually part of his online persona … he is not a corporate blogger. My take is that he wanted to talk about what he was seeing, thinkingVariationen von poker regeln. and hearing in sections of the HPC market. Well worth the read, and as always, Doug is insightful and incisive. Adding it to my blogroll.

On Sunday 2-March-2008, 18TB of data was moved 1050 km in 12 hours. The network fabric and technology that brought this to being? The US interstate system, and our truck. Physical transport of media is still the bandwidth leader. This means that the interstate system transported about 1.5 TB/hour, or 0.42 GB/s. 420 MB/s And yes, a JackRabbit did make its way across Ohio, West Virginia , Virginia, and finally came to rest in its new home in North Carolina.

Componentized, stripped of all garbage^H^H^H^H^H^H^H^Hthings you don’t need in a server, modularized … Ok, so when can we play with it to see if it takes to our JackRabbits and clusters? We are loading almost all our OSes via diskless/CF, and I would love to do this with WS2k8. That and I want to get iozone, bonnie, and other tools ported. Our io-bm should work nicely, all we need is an MPI stack, and we can do parallel IO.

Readthis, this morning on /. . In short Microsoft will implement its own GNU compatible environment. Why is it almost a good thing? Simple. There exists a great environment now, for all of this. Called Cygwin. I had been trying to convince the Microsoft people for a while now, to get behind this effort, and support this wholeheartedly, on windows. I made the point to Kyril Faenov at SC07, and to multiple others at Microsoft for the past 2+ years.

Well, we did it. We now have 4 GB CF images to boot our JackRabbit’s from SuSE 10.2 and 10.3, RHEL 4 and 5, OpenFiler 2.2 (2.3 also works), and Ubuntu. This is nice in that it is simple to replicate our installs. Installing SuSE 10.2 is painful (Zenworks … thats all you have to say) Myricom 10 GbE drivers, and we will make sure we have the Intel ixgbe drivers as well.

For a while now, privately, and publically, I have been suggesting to the good folks at AMD that they ought to build an 8 core chip, literally by gluing two quad Barcelona’s to a die and connecting them with Hypertransport. The point I have been making is that Intel is going to do something like this, really soon, and if they wish to compete, they ought to get to market first with their version.

/. linked to this story on cio.com. I read it, and there are some gems. Things I have seen corporate IT leadership folks do. While reading it, I was thinking “gee, wouldn’t it be funny if the problem of vendor favoritism showed up?”. That is, when specific vendors are chosen above others, not because of technological reasons, or valid business reasons, but because the CIO or IT leader wants to do business with people they know.