Jim Larus from Microsoft Research gave a nice overview of their research in the large datacenter space. He pointed out the problem of limited bisection bandwidth and how that prevents applications from running across clusters. Microsoft is working on a mesh network called Monsoon connecting adjacent datacenters. Besides the mesh network, two of the key ideas in Monsoon are Valiant Load Balancing and source routing. Source routing keeps the switches simple and largely stateless and avoid hot spots. Mike Foss from Rice who presented a poster later also was working on something similar in principle that employed source routing to its advantage.
Another problem that Jim mentioned of which I heard about in several other peoples' research was putting idle machines into a low energy sleep state in order to conserve power. There are interesting challenges here. For instance, a network packet or a timer interrupt is sufficient for "disturbing" the sleep of conventional operating systems. Tickless linux solves part of the problem. For the networking part of the problem, someone is working on a microcontroller that can address most of the stuff without involving the processor. Others were working on disabling ethernet ports.
A direction MSR is exploring is "self-aware" and "self-managing" grid applications which automatically reconfigure themselves based on input and SLA requirements. It was great to hear that Jim's team is thinking about this problem since I've been thinking about this for a while myself and I should read up more on MSR's publications in this area. Ben Sigelman from Google pointed out later in the session how this is a hard problem in general and there are less powerful but practical alternatives (like visualization).
Also, Jim was very optimistic about PCM (Phase Change Memory) as a more scalable and efficient alternative to FLASH. PRAM is quite expensive today but Jim felt it was a matter of time before it was cost effective. Chuck Thacker from MSR also later showed some confidence in PCM but did not quite share Jim's enthusiasm.
This also came up in the next talk by Sarah Bird, who is a grad student at UCB about hardware performance counters. Sarah's research deals with providing a single consistent view across all the many performance counters in a multicore machine + adding many multicore specific counters to the hardware -- and doing this all in a standard way across platforms. Ben Sigelman pointed out that integrating this information with a distributed tracing service might be useful. Sarah also talked about the Autotuning project they are working on at Berkeley. Several people in the audience pointed out that adding new hardware support is a bad idea. But Sarah pointed out to a previous project (PAPI) that aimed to solve the same problem in software was a failure.
This was followed the poster session. The next speaker was Richard Kaufman from HP Labs. Richard's point of view was that of a server supplier as compared to MSR's make-your-own-servers. So he cited several general purpose trends in power and datacenter design.
The final speaker was the Chuck Thacker. He presented the case for modular datacenters based on containers, several of which will be deployed in a "parking lot" type datacenter. He gave a good high level overview of the technology required, including left to right cooling, network switches, and rack space considerations. In response to an audience question, he seemed to be pretty determined not to violate the absolute symmetry in the system -- all containers are meant to be equal, so you cannot have special containers with high level network switches for example. I found a nice overview of Chuck's talk here.
