scripts/lbnl_hw.nhc: nhc_hw_gather_data is slow at parsing /proc/cpuinfo

[sdiak]

Hello,

on a cluster that I have installed for a client, the function nhc_hw_gather_data() spends a lot of time parsing /proc/cpuinfo.

[root@<redacted> ~]# time nhc

real    0m9.824s
user    0m1.532s
sys     0m8.232s

With the proposed patch, the time is now reduced by a factor of 5+

[root@<redacted>~]# time nhc

real    0m1.692s
user    0m1.563s
sys     0m0.105s

The patch works by caching the content of /proc/cpuinfo in a local variable

Informations :

• OS : Red Hat Enterprise Linux release 8.8 (Ootpa)
• Kernel : 4.18.0-477.27.1.el8_8.x86_64
• CPU : Dual-socket with AMD EPYC 9654 96-Core Processor

0001-Makes-proc-cpuinfo-parsing-faster-by-caching-the-con.patch

[OleHolmNielsen]

We have similar nodes, and NHC is somewhat faster:

time nhc

real 0m3.018s
user 0m0.253s
sys 0m2.084s

OS: Rocky Linux 8.9 (Green Obsidian)
Kernel: 4.18.0-513.18.1.el8_9.x86_64
CPU: Dual-socket AMD EPYC 9474F 48-Core Processor

[sdiak]

This is running on a IDLE node where CPUs are set to powersave

[sdiak]

This is a problem on this given site because CPU throttling when a power supply is lost makes slurm draining the node :

• without the patch : nhc times out and slurm drains the node
• with the patch : time nhc is 19.992s and the node stays in production with degraded performances