History

Noted (if not noteworthy) events or activities taking place over the course of the project.

open cabinet showing nine blades from Build-a-Blade and gigabit switch

12/8/2014: A new 45-node blade-format cluster has been running continuously for the last week. We decided to purchase from Build-a-Blade and we have been very pleased with the quality of the product and the customer service that they have provided. (Plus, I love the blue color!) Preliminary tests suggest the new cluster is more than ten-times faster than our previous one.

8/24/2013: Over the last 14 weeks, we have lost at least six disk drives in the compute nodes. We are down to a few spare drives.

5/10/2013: After an unsuccessful attempt during the Fall 2011 semester following the disassembly and transfer to our new lab in Winter Hall, two research students successfully reconfigured the cluster with Rocks 6.1 Emerald Boa.

9/9/2008: The cluster's age is starting to show. This summer, we had periodic problems with compute nodes refusing to reboot. A few day ago, one node completely died. Also, our brand new switch has developed a bad port.

7/2/2008: After replacing the gigabit switch, I clocked 83.75 GFLOPS on the HPLinpack.

4/2008-6/2008: With several students from my machine learning class, we implemented and tested a genetic algorithm using the island model. The algorithm was tested on the chess endgame domain from the UCI machine learning repository. Our resulting paper appears in the Proceedings of the 2008 International Conference on Genetic and Evolutionary Methods.

6/25/2007: Have been running MĘDEN simulation experiments on the cluster as part of our service research. We are currently analyzing 460K trials that we have completed. (We've probably run three-times that but had to throw the results out for various reasons.)

Our second cluster (minus two nodes)

7/2006: Benchmarking the cluster using the HPLinpack, we clocked 74.1 GFLOPs.

5/2006: The school installed a dedicated air-conditioner for the cluster. A ceiling fan works much of the time when it is nice outside and we use the air-conditioner when it gets too hot.

8/31/2005: Much has been happening even though this hasn't been updated. Thomas Cantrell's parallel search code progressed to the point where we were able to cover the search space of the ClimbPro12 puzzle. Various obstacles prevented a solution to ClimbPro24.

1/4/2005: Over three months now and we are still waiting on the last two compute nodes.

12/8/2004: Ten weeks and counting from the placement (and payment) of the order for our new 32-node cluster. The disks for the server apparently arrived; we're just waiting on two missing shuttles. I've had all but eight of the nodes off during the last few days as the heat was getting quite high. We're still working with facilities folks here to get a fan to vent the heat from the cluster. Thomas Cantrel is making progress on an implementation of a parallel depth-first search.

12/3/2004: After nine weeks we have 30 out of 32 compute nodes and today we received the gigabit switch. I must admit that some of my faith in corporate service -- at least for CDW -- has been restored. There is no avoiding the fact that this has been a nightmare, but the representative from CDW seems to have gone the extra mile and I anticipate being satisfied at the end of the ordeal. I have mildly amended my comments from 11/20/04; they were written at the peak of my frustration.

11/20/2004: More than eight weeks after ordering our new cluster, we still have nothing to show but frustration! CDW sent us the wrong items, they sent them unassembled even though we paid for assembly, they told us things were in stock when in fact they are back ordered, they told us it would ship "today" or "tomorrow" at least four times already! You would think that with an order of many tens of thousands of dollars, they would give it their best effort. After a quick web search, I find that I am not alone in my poor experience with CDW. Probably like any corporate service encounter, it depends greatly on the person one is actually working through.

6/8/2004: We've written and run some simple parallel programs using the MPI libraries. (Fun stuff!)

4/6/2004: All nodes are configured. We benchmarked the between-node communication bandwidth in April or May. Results were basically what you would expect for a cluster based on 100Mb ethernet.

3/8/2004: With the change in the weather, I had shut down the cluster, but now that our air conditioning is working again, we'll plan to keep the frontend and two compute nodes (compute-0-0 and compute-0-1) up by default.

2/19/2004: We've had the frontend configured for some months, but we finally have 12 nodes configured. Four remain to be installed and configured and then initial benchmarking can begin.

Original 'GooGoo' cluster pre-retirement after removal from rack

2/5/2004: We learned today that we received a generous grant that (among other things) will fund the construction of a 32-node cluster using current CPU speeds.

11/19/2003: We have made some visible progress. Kirsten Iba started helping and we have the 16 compute nodes physically configured and four have been installed.

9/22/2003: We have received a generous donation of additional equipment that will bring this prototype to a 16-node cluster with an extra master. We should be starting assembly during the coming week or two.

9/18/2003: We have obtained a collection of donated equipment that should allow us to build a prototype 11-node cluster using channel bonding over a 100Mb/s ethernet backplane. We are currently gathering information on how we can network boot and configure the nodes in the cluster, on the advantages and disadvantages of alternate specific configurations, and about whether we can easily "dual-boot" the cluster into more than one configuration.