I am using Julia v0.5.0 on a memory limited SGE cluster. In particular,  to 
submit jobs on the cluster, both h_vmem and tmem resource flags need to be 
passed to in qsub command. However, all of Julia jobs keep on being killed 
because workers seem to be very hungry for virtual memory and ask for it 
outside of the SGE scheduler. 

A more detailed description of the problem can be found in the last post in 
this thread 
It seems that the best workaround (listed on this thread 
<https://github.com/JuliaLang/julia/issues/18477>) is to change  #define 
REGION_PG_COUNT 16*8*4096 to #define REGION_PG_COUNT 8*4096 in 
src/gc-pages.c when compiling Julia. 

However, I am just a user of the cluster and don't have permission to 
compile Julia on it. Nor do i have access to machinery to compile my own 
binary . As such I am wondering if there could be a more user friendly 

Two ideas come to mind:

1) Offer a low virtual memory for Julia binaries; or
2) Add a flag when opening julia flag that lets you set K where K is 
defined as REGION_PG_COUNT K*8*4096 ... or some other equivalent...

In particular, option 2 would really be addition to the language and 
greatly aid users - especially those without capacity to compile julia. If, 
for some reason, this is undesirable, perhaps option 1 would present the 
best short term fix?

Reply via email to