[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [HTCondor-users] A GPU server configuration using HTCondor

oops. I left out this part of the configâ


num_slots_type_1 = 1

num_slots_type_2 = 1




From: John M Knoeller
Sent: Friday, June 8, 2018 6:06 PM
To: htcondor-users@xxxxxxxxxxx
Subject: RE: [HTCondor-users] A GPU server configuration using HTCondor


Cpus=0 is effectively shorthand for Cpus=all

there is no way to assign a slot to have 0 cpus, since there is no way to run a program that uses doesnât use the CPU.

Also, HTcondor currently doesnât support fractional cpus.  so the minimum you can assign is cpus=1


If you want to have one slot with a GPU and minimal, cpu and

another slot with all of the cpus and no GPU, then you would do something like this.


cpus = $(DETECTED_CPUS)+1


# this gets 1 âbonusâ cpu, since we declared cpus to be detected+1

# we expect the cpu usage to be minimal.

slot_type_1_partitionable = false

slot_type_1 = cpus=1, mem=20000, gpus=1


# This will get the remainder after 1 cpu and 20000 memory are carved off.

slot_type_2_partitionable = true


From: HTCondor-users [mailto:htcondor-users-bounces@xxxxxxxxxxx] On Behalf Of Ali Rajabi
Sent: Thursday, June 7, 2018 8:15 PM
To: htcondor-users@xxxxxxxxxxx
Subject: [HTCondor-users] A GPU server configuration using HTCondor


Hi everyone,


I am trying to configure HTCondor to be able to submit jobs to a GPU server. The jobs will only need to run on GPUs, but the way I have defined each slots, all CPU threads are also available and Condor distribute the process on all CPUs as well as GPUs. 


I have 6 GPUs and defined 6 different slots as:


slot_type_1_partitionable = false

slot_type_1 = cpus=0, mem=20000, gpus=1

num_slots_type_1 = 1


It looks like Condor overrides the "cpus=0" part and make all 32 CPU threads available to the jobs I submit.


I would appreciate if anyone could help me make Condor ignore CPUs all together.


Best Regards,