Cray User Group: Day Three
May. 11th, 2016 11:14 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
sawylUp early for the system monitoring SIG at 07:30am. Rather to my surprise both C and A were there despite yesterday's late night. After scoffing down a load of pastries, I escaped for the morning — yet another plenary — only to get caught in rush hour traffic, increased security arrangements, and catastrophically heavy rain. Realising that I didn't need to be back for the first afternoon session, I waited out the rain and grabbed lunch on the way back.
The first couple of sessions were on power management and power facilities in Knight's Landing. The first was on Sandia's PowerAPI while the second was a retread of information presented in Monday morning's tutorial — which just happened to be given by the same speaker!
I then sat in on a NERSC talk about their migration from Grid Engine to SLURM. There were some interning points about native SLURM — particularly the way it allowed them to run very large numbers of serial jobs on the same nodes — and some of the problems they'd observed. The CSCS paper on their production Cray CS system was interesting, but the amount of work they'd had to do to get it working sounded scary.
The most interesting paper of he afternoon was the most unexpected: a comprehensive analysis of the Gemini interconnect on Blue Waters, which concluded that even successful link recovery events were sufficient to cause jobs to fail. It was a fascinating insight into the sort of odd corner cases that can develop on a very large system and somewhat sobering to realise that the sorts of system tasks we normally assumed were neutral actually had an impact on the work and could cause things to crash.
In the evening, LA and I eschewed the organised Routemasters — I remember when they were used in earnest — and walked to the museum of London via St Pauls. As we were leaving someone mistook me for someone from Germany — doubly funny because LA had made the same mistake yesterday at lunch!
We got to the museum on time, wandered around with nibbles and drinks before sitting down to a rather nice three course meal. Over supper CM convinced me that we needed to set up a centre of high performance computing excellence and that I should use be its PhD poster child. I gave him official permission to bully me into doing it.
Rather than go on to a pub, I returned to the hotel with TL and the others planning to get up super early for Bill's PBS BOF at 07:30 tomorrow...
The first couple of sessions were on power management and power facilities in Knight's Landing. The first was on Sandia's PowerAPI while the second was a retread of information presented in Monday morning's tutorial — which just happened to be given by the same speaker!
I then sat in on a NERSC talk about their migration from Grid Engine to SLURM. There were some interning points about native SLURM — particularly the way it allowed them to run very large numbers of serial jobs on the same nodes — and some of the problems they'd observed. The CSCS paper on their production Cray CS system was interesting, but the amount of work they'd had to do to get it working sounded scary.
The most interesting paper of he afternoon was the most unexpected: a comprehensive analysis of the Gemini interconnect on Blue Waters, which concluded that even successful link recovery events were sufficient to cause jobs to fail. It was a fascinating insight into the sort of odd corner cases that can develop on a very large system and somewhat sobering to realise that the sorts of system tasks we normally assumed were neutral actually had an impact on the work and could cause things to crash.
In the evening, LA and I eschewed the organised Routemasters — I remember when they were used in earnest — and walked to the museum of London via St Pauls. As we were leaving someone mistook me for someone from Germany — doubly funny because LA had made the same mistake yesterday at lunch!
We got to the museum on time, wandered around with nibbles and drinks before sitting down to a rather nice three course meal. Over supper CM convinced me that we needed to set up a centre of high performance computing excellence and that I should use be its PhD poster child. I gave him official permission to bully me into doing it.
Rather than go on to a pub, I returned to the hotel with TL and the others planning to get up super early for Bill's PBS BOF at 07:30 tomorrow...
