Temp Dir Selection

[ahansen1]

I am trying to understand the algorithm for temp dir selection. In the below capture it used temp ssd3 back to back when in theory ssd2 was available to use. I had also seen a similar case earlier when it selected ssd1, ssd2, ssd1. Is this behaving as expected? I was concerned that it was putting too much load re-using the same dirs again too soon.

   #    plot id    k    tmp     dst   wall   phase    tmp     pid   stat    mem   user    sys     io                   
 0   ae6dcf38   32   ssd3   buff2    46s     1:1    15G   93938    SLP   343.0M    48s    17s     0s
 1   3608382c   32   ssd3   buff1   0:36     1:3   113G   91687    SLP     4.2G   0:49   0:02     1s
 2   9b8b2181   32   ssd1   buff2   1:12     1:4   157G   89626    SLP     4.2G   1:32   0:03     5s
 3   f7367686   32   ssd2   buff1   1:57     1:5   174G   86960    SLP     4.2G   2:28   0:05     4s
 4   5efc3477   32   ssd1   buff2   2:33     1:6   179G   84788    SLP     4.2G   3:09   0:06     7s
 5   316ba756   32   ssd2   buff1   3:09     1:7   170G   82529    SLP     4.1G   3:58   0:08     7s
 6   ae76ff23   32   ssd3   buff2   3:45     2:2   206G   80212    RUN     1.5G   4:31   0:09   0:02
 7   db9a96d7   32   ssd3   buff1   4:23     2:5   222G   77877    RUN     1.3G   5:04   0:11   0:05
 8   19d9b478   32   ssd1   buff2   5:02     3:2   225G   75461    RUN     4.0G   5:34   0:12   0:10
 9   818f9dc4   32   ssd2   buff1   6:09     3:4   166G   71344    RUN     4.0G   6:34   0:15   0:16
 10   d9351d55   32   ssd1   buff2   6:45     3:5   155G   69073    RUN     4.0G   7:00   0:16   0:23

[ahansen1]

I don't know python at all but was trying to dig into the code. Is this only using the major or minor phase of the job to determine the sort and then grabbing the first returned when determining which temp directory to use? When using more than 2 temp directories I'm wondering if this is not adequate? I could be completely misreading this code.

        # Plot to oldest tmpdir.
        tmpdir = max(rankable, key=operator.itemgetter(1))[0]

[ahansen1]

I was digging into the code further trying to understand.

rankable = [ (d, phases[0]) if phases else (d, job.Phase(known=False))                for (d, phases) in eligible ]

If I am reading this correctly it's grabbing the first (earliest) phase for from the sorted phases for each directory. So I think rankable becomes a tuple of the directory and it's earliest phase, e.g. for our example (ssd2, 1:5), (ssd3, 1:3). I think the itemgetter(1) that I saw above is then grabbing the phase object not just a property of the phase. Though, this interpretation of the code seems to indicate it should have selected ssd2 which would have a higher phase, but instead it choose ssd3 again.

Is anyone else seeing troubling scheduling when using more than 2 temp drives?