Hi all,
I have done some experiments on parallel kernel dumping. I would like to
share the test result with you. Hope it helps.
Test environment:
Machine: HP ProLiant DL980 G7 with 4TB RAM.
CPU: Intel(R) Xeon(R) CPU E7- 2860 @ 2.27GHz (8 sockets, 10 cores) (4
CPU were enabled the 2nd kernel by nr_cpus=4)
Kernel 3.9.0-rc7
kexec-tools 2.0.4
makedumpfile v1.5.3 with lzo library
crashkernel=4096M (I have tested with 2048M but failed with OOM on 3 or
4 parallels dumping in cyclic mode)
I didn't get a real multipath storage device, so I just put dump files
on 4 different disks via 3 HP Smart Array controllers. (mounted on /0,
/1, /2 and /3 in the capture kernel)
Measured time like this (for example: lzo compression, non-cyclic, 4
parallels):
time makedumpfile -l -non-cyclic --split --message-level 23 -d 31
/proc/vmcore /0/vmcore_0 /1/vmcore_1 /2/vmcore_2 /3/vmcore_3
I run several tests with different option, parallels from 1 to 4, and
combined with zlib and lzo compression.
Test result:
-----------------------------------------------------------------
| |Parallels 1|Parallels 2|Parallels 3|Parallels 4|
-----------------------------------------------------------------
|zlib cyclic | 42m25.321s| 34m0.168s| 29m44.908s| 28m50.387s|
-----------------------------------------------------------------
|zlib non-cyclic| 42m7.842s| 28m28.275s| 23m25.750s| 21m6.476s|
-----------------------------------------------------------------
|lzo cyclic | 23m40.010s| 18m19.932s| 21m47.903s| 22m47.605s|
-----------------------------------------------------------------
|lzo non-cyclic | 20m45.749s| 16m42.045s| 15m41.070s| 15m18.605s|
-----------------------------------------------------------------
--
Thanks,
Jingbai Ma