Routes and analytics
Jul. 14th, 2016 09:58 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
sawylAfter an early night yesterday, as recommended by my medical professional, I felt a lot better despite being woken pre-dawn by the local herring gulls — presumably getting excited over abandoned kebabs. Fortunately, getting up and closing the window shut out the racket and I managed to fit some more sleep in before the alarm went off.
The morning was taken up with an interesting problem: determining the levels of backfill in PBS, matching both placed jobs and backfilled jobs with projects, and attempting to show whether one group was being backfilled more than another. Combining scheduler log data with start records in the accounting data, I was able to assemble enough data for each project to determine the raw job counts, the number of nodes, and the number of node seconds for both placed and backfilled jobs. These confirmed that there had been a period when a small number of groups had benefited from backfill, but without more information about the queues it was hard to tell exactly why the situation changed partway through the data.
I now wonder whether there might be some more interesting ways to use this data. I think it might be useful to maintain the order of jobs in the priority list, so when each backfilled job is correlated with a project, it is possible to determine whether the job has been run ahead of other, higher priority candidates. This might be enough to report on the level of FIFO-ness in the queue and thus, the impact of priority ordering on the overall throughput of work for particular project groups.
With work over, I went down to the quay to meet up with L&A for a few routes. Having assured me that they were going to be on time because A was driving — something L said was code for only being 10 minutes late! — they were actually there well ahead of time and had to wait for me to get changed. We climbed some fun stuff, including a couple of difficult 6b+ TR routes which I comprehensively mis-read and consequently failed to flash.
After the others had tired a bit, we went into the boulder rooms and did a few problems. I successfully repeated the dyno problem I'd finished at the weekend — I'm now confident I've got it nailed — and the others tried the first couple of moves on the 40 degree stamina circuit. They did pretty well considering they were climbing front-on — I can only manage somewhere between 20-30 moves on the easiest circuit if I go full on with knee drops — but we soon called it night, not least because having all that weight on rough, juggy holds does horrible things to your hands in next to no time.
The morning was taken up with an interesting problem: determining the levels of backfill in PBS, matching both placed jobs and backfilled jobs with projects, and attempting to show whether one group was being backfilled more than another. Combining scheduler log data with start records in the accounting data, I was able to assemble enough data for each project to determine the raw job counts, the number of nodes, and the number of node seconds for both placed and backfilled jobs. These confirmed that there had been a period when a small number of groups had benefited from backfill, but without more information about the queues it was hard to tell exactly why the situation changed partway through the data.
I now wonder whether there might be some more interesting ways to use this data. I think it might be useful to maintain the order of jobs in the priority list, so when each backfilled job is correlated with a project, it is possible to determine whether the job has been run ahead of other, higher priority candidates. This might be enough to report on the level of FIFO-ness in the queue and thus, the impact of priority ordering on the overall throughput of work for particular project groups.
With work over, I went down to the quay to meet up with L&A for a few routes. Having assured me that they were going to be on time because A was driving — something L said was code for only being 10 minutes late! — they were actually there well ahead of time and had to wait for me to get changed. We climbed some fun stuff, including a couple of difficult 6b+ TR routes which I comprehensively mis-read and consequently failed to flash.
After the others had tired a bit, we went into the boulder rooms and did a few problems. I successfully repeated the dyno problem I'd finished at the weekend — I'm now confident I've got it nailed — and the others tried the first couple of moves on the 40 degree stamina circuit. They did pretty well considering they were climbing front-on — I can only manage somewhere between 20-30 moves on the easiest circuit if I go full on with knee drops — but we soon called it night, not least because having all that weight on rough, juggy holds does horrible things to your hands in next to no time.
