CHOLMOD slows down in parallel
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
suitesparse (Ubuntu) |
Confirmed
|
Undecided
|
Unassigned |
Bug Description
If I run a simple CHOLMOD C++ program, I noticed that it takes more time running in parallel than single threaded. For the attached program file I did
g++ -o main main.cpp /usr/lib/
time ./main
-> real 2m21,367s user 11m30,913s sys 5m11,448s
(now I force it to be single threaded)
time OPENBLAS_
-> real 0m49,684s user 0m48,922s sys 0m0,760s
Notice the huge "sys" time consumption in parallel run.
I have a 4 core / 8 thread Intel CPU running Ubuntu 18.04, everything up-to-date.
In my research project with a more complex matrix, the slow-down is the factor 8.
Further info: It is not machine-dependent. It is also confirmed in Debian 10 on several different machines.
Status changed to 'Confirmed' because the bug affects multiple users.