I’m on my way back to Europe from the Symposium on Operating Systems Principles (SOSP), held in a stunning place in Montana. Wow, it was a huge event: over 500 attendees, good talks, lots of interactions and tones of food for thought to chew on. Many famous researchers rubbing shoulders with wanna-be OS writers, bright PhD students and quite many industry representatives (including myself) fishing for new ideas (Amazon, Facebook, Nokia even Apple were represented).
So, what would be the key takeaways?
There were two papers – one from CMU (even won the best paper award) and one from Microsoft research in co-operation with Johns Hopkins University and Washington University – both reaching the same conclusion: same computing throughput can be obtained in data centres using arrays of low(er) capability chips (such as Intel Atom) than with server chips, but at roughy 1/10 of the power budget. This was quite a revelation for me and certainly something to think about further.
There were then several papers and posters on novel OSes for massively multi-core chips, mostly sponsored by – surprise… – Microsoft. There seems to be a consensus building up around the fact that today’s time-shared OSes (Linux, Windows, MacOS) will not scale and the most promising way forward is the use of space-shared OS concepts (it was nice to see my thoughts confirmed…) Microsoft is also making forrays into OS for heterogeneous chips, covering e.g. GPUs.
A slightly worrying result – from a couple of papers – was that even with state-of-the-art technologies reliable recording and replay of parallel software behaviour is either unacceptably intrusive (overhead of 50-150%) or so unreliable that it becomes unusable in practice. This clearly is an area where would be great to see some breakthrough results…
Beside these major themes, I thoroughly enjoyed some of the sessions, such as Prof. Barbara Liskov‘s keynote (last year’s ACM Turing Award winner), the panel at the HotPower workshop or VMWare’s BoF session. Prof. Liskov re-iterated her belief that programming centred around abstract data types and encapsulation is the way to go and expressed her reservations with regards to how far chips will scale in terms of number of cores. This resonated quite well with Flautner Krisztián’s (ARM) thoughts at the panel; he was also quite clear in his message that only radical – 40% and up – power reduction really counts, anything less is a waste of research effort. Point taken… It was good to hear him stating that he believes that eventually homogeneous chips will prevail, even in the embedded / mobile space.
VMWare hosted a really interesting session primarily aimed at getting universities interested in working with them. Somewhat surprisingly, the guys – real engineers, no marketing bullhead was present – gave the impression that Amazon is VMWare’s main competitor (not a word about MS or Citrix). They were quite clear that they believe in open cloud computing standards and will push for it, primarily in DMTF. The research issues they raised were mostly related to scaling up to 1000 data-centre nodes and especially efficient resource management (I/O was a major concern), VM re-allocation / load balancing algorithms. VMWare gave the impression of an open company eager to work with academia on really exciting research issues.
Finally, here are a few gossip subjects / stuff I picked up:
- Google uses the same infrastructure for public services (such as the search engine) and internal development work; instead of power management, employees shall use the farm during low activity hours
- the search engine and indexing are fully distributed with no or very little co-ordination, thus the same search done at the same time in the US and Europe may give two completely different results
- Steve Ballmer indeed simulated destroying an employee’s iPhone at a Microsoft all-staff meeting (confirmed by two MS guys); at Apple this would never happen – anyone caught with anything else than iPhone would be fired by Steve Jobs, immediately, especially since everyone was given a free iPhone when it was launched (says a former Apple employee). Now you know