A Blog

Today was the final day for Team Purdue in the Cluster Challenge. The day broke both early and late.. My teammate, Paul, and I stayed late and worked on our machine to rerun the HPCC benchmarks with a “no holds” on the power budget except for the hard limit of the power circuits themselves. Our runs, after performing various unnatural acts to our machine from SiCortex, got us 1.07TeraFlops. Was the stress and lack of sleep worth it? Yuup, we were the first team to break the teraflop level at the Cluster Challenge. Taiwan would eventually break our new record with around 1.1TF, but we had other plans for our machine that prevented us and the machine from achieving the highest score possible…

The day after the runs complete in the Challenge, that day is used for team’s to show off to the crowds and focus on high performance computing in education. With much planning from SiCortex, Purdue helped break out the bike generators and a mess of DC power gear and had a racing team come in. They were successfully able to power our machine long enough for us to recreate a water simulation that took a team of engineers three years to do by hand. It was quite amazing.

So, if staying late to break the 1TF wasn’t bad enough, getting up to do the bike demo was certainly trying. After some waiting, the awards ceremony took place… Dresden/IU won. I think I’ll let that topic drop.

Packing up the machine was entertaining. Since the machine was getting hauled from Texas to Massachusetts, everything had to be removed and boxed safely. We got all the modules out and in boxes, pulled a backup of our data from the service processor, and sent the machine away. It appears that the team performed so well that we were invited to the SiCortex booth on the show floor to help with their tear down. Normally, this would just be a lot of work without much award. However, I got to help disassemble a fully loaded SC5832. Man, there are just too many boards in that thing. All the same though, I can’t recall a time when more money moved through my hands than when unloading that beast.

This pretty much concludes SC08.. Tomorrow afternoon I begin traveling home from another great year.

This post is a little late, but on day four of SC08, the second Cluster Challenge finished up. The data sets were mightily tricky and seemed to favor very large memory machines with just some cores. Nonetheless, our team did well. We certainly got all the applications spinning and kept the machine at full utilization for nearly every minute of the competition. Not a single hardware problem during the competition time, I’m very happy about that.

After the deadline came and past, several of the teams threw caution to the wind and cranked their clusters up. The coordinators wanted to see someone get a Linpack score of 1TF, and we were glad to push our machine to the limit to archive it. With the great help of Rich from SiCortex, we really got the machine going: 1.07TF. Definitely not the percent of peak we could have gotten, but for a late night run through the dark, we were proud.

I finally got back to the hotel this morning and got some sleep before heading back to the floor at noon local time. The cluster and the guys were still there and everything was running smoothly. Jobs continue to progress through the queue and no real hardware failures yet!

This afternoon, we all went to the Top 500 Birds of a Feather session and found out about the fastest machines in the world. We even got a mention during the session, whee! Then, it was back to check on the machine and then to a party thrown by Platform Computing and Intel.. Hopefully, things go well tomorrow on the clustering front. The challenge ends, so there’ll be an update tomorrow about how several month’s of work turned out.

Today was the official start of the 2008 Cluster Challenge. At 8AM CST the teams went onto metered power and the HPCC running begun. Purdue arrived a little after eight and begun running the benchmark. We got a number, submitted it. Then, we attempted for a better number… That’s when we found out the results coming out of the benchmark were not correct. After some hardware replacement (first time Purdue has had to replace a module from a SiCortex), things were back in action. As time ran to a close, we submitted our final numbers.

Little did we know that the Taiwan team would come to score 9Gflops more than us to take the largest number of flops..

After some fan fair and picturing taking, the teams were off with the data sets. As of about 2AM, things are looking to be coming right along. Just a few members of the team are staying overnight to babysit the box as it runs through the night.

Today the Purdue Cluster Challenge team flew in from Indianapolis to Austin, Texas. After getting to the convention center a after 2PM, we met up with the SiCortexian guy who helped get our machine and get it all set up. Thankfully, most of the prep work was done (though there aren’t really any heroics involved in installing one) and it more or less just rolled off the pallet. We had a small networking mis-configuration issue and forgot to pack a monitor and keyboard… Some help from the regular Purdue booth and a couple minutes reconfiguring Gentoo, and the machine was running HPCC

We got to finally see the hardware the other teams were using. Some teams (like Alberta) brought more of the same from last year. Others brought snazzy visualization things and one team brought a blade-center like tower. MIT has yet to show up… Rumor has it that they lost their “green” powered ride and had to fly down. Poor guys!

It appears our machine made it safe and sound, and now it’s time to watch Stargate before getting to bed early