Nikhef High Mem Observations - PowerPoint PPT Presentation

jolene
nikhef high mem observations n.
Skip this Video
Loading SlideShow in 5 Seconds..
Nikhef High Mem Observations PowerPoint Presentation
Download Presentation
Nikhef High Mem Observations

play fullscreen
1 / 7
Download Presentation
Nikhef High Mem Observations
79 Views
Download Presentation

Nikhef High Mem Observations

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript

  1. Nikhef High Mem Observations Jeff Templon GDB 2014.02.12

  2. Which High Mem Problem? • Normal LHC jobs are somewhat high mem … Nikhef limit is 4096mb pvmem Jeff Templon, Multicore TF #2

  3. PVMEM • Pvmem limit? • Translates to ulimit on process • Process gets out-of-mem signal instead of kill • Chance to handle • Room for overhead outside of app space • We set this at queue level … job default, apps don’t have to do anything. Jeff Templon, Multicore TF #2

  4. What if you need more? • Can ask for it • Most ways that should work, don’t • Dennis van Dok's investigations • This does “work”: • CERequirements = "other.GlueHostMainMemoryRAMSize >= 8192” • However, not consistent … Jeff Templon, Multicore TF #2

  5. Inconsistency Resource_List.mem= 8192mb Resource_List.neednodes= 1:ppn=5 Resource_List.nodect= 1 Resource_List.nodes= 1:ppn=5 Resource_List.pvmem= 2936012800b • Pvmemlimis 2800mb from ‘infra’ queue (dteam) data seg size (kbytes, -d) 8388608 max memory size (kbytes, -m) 8388608 cputime (seconds, -t) unlimited virtual memory (kbytes, -v) 8398848 Jeff Templon, Multicore TF #2

  6. Link between mem & cores • Current : ask for “mem” (not vmem) then get (mem / 2GB) cores • No longer so relevant … • 2 / 3 / 4 / 8 GB phys per node in service • Maui is capable of steering high mem jobs to right nodes • Could accept up to 8G / core as long as these jobs are a minority • Currently reviewing situation Jeff Templon, Multicore TF #2

  7. BDJ vs many processes • Maui doesn’t understand 1st one • BDJ needs 32 GB and 8 cores • Maui looks for one node with 256 GB and 8 cores freed (32 GB per core) • Look for info : mostly find other people with same questions • Not clear at this point whether it’s worth source spelunking to figure it out. Jeff Templon, Multicore TF #2