Re: Wait LCP to ensure durability

[Thomas Waibel-BGo]

We are now busy restoring diskbased tables and seem to experience a CPU bound situation.

We have on each Node 24 CPUs (2x Intel(R) Xeon(R) CPU E5-2643 v2 @ 3.50GHz with each 6 cores and 12 threads). Using Linux cmdline isolcpus=7-23 we assign 16 CPU threads to NDB Cluster. The remaining 8 CPUs are used for OS and one MySQL API node.
...
ThreadConfig=ldm={count=12,cpubind=12,13,14,15,16,17,18,19,20,21,22,23},tc={count=1,cpubind=7},recv={count=2,cpubind=8,9},main={count=1,cpubind=10},io={count=1,cpubind=11}
...
Picture in nmon looks like this:
┌nmon─14g──────[H for help]───Hostname=ndb13────────Refresh= 2secs ───14:54.59────────────────────────────────────────────────────────────────────
│ CPU Utilisation ─────────────────────────────────────────────────────────────────────────────── │
│---------------------------+-------------------------------------------------+ │
│CPU User% Sys% Wait% Idle|0 |25 |50 |75 100| │
│ 1 3.9 3.9 0.0 92.3|Us > | │
│ 2 12.5 12.5 0.0 75.0|UUUUUUssssss > | │
│ 3 3.5 6.9 0.0 89.6|Usss > | │
│ 4 6.5 7.0 0.0 86.4|UUUsss > | │
│ 5 3.5 5.1 0.0 91.4|Uss > | │
│ 6 12.6 8.1 0.0 79.3|UUUUUUssss > | │
│ 7 4.5 6.9 0.0 88.6|UUsss > | │
│ 8 16.1 14.0 0.0 69.9|UUUUUUUUssssss > | │
│ 9 6.6 9.6 0.0 83.8|UUUssss > | │
│ 10 3.5 7.5 0.0 88.9|Usss > | │
│ 11 0.5 0.0 0.0 99.5| > | │
│ 12 95.0 5.0 0.0 0.0|UUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUss> │
│ 13 12.8 10.3 0.0 76.9|UUUUUUsssss > | │
│ 14 6.8 8.9 0.0 84.3|UUUssss > | │
│ 15 9.2 10.3 0.0 80.5|UUUUsssss > | │
│ 16 6.0 10.6 0.0 83.4|UUUsssss > | │
│ 17 10.6 9.0 0.0 80.3|UUUUUssss > | │
│ 18 11.0 10.0 0.0 79.0|UUUUUsssss > | │
│ 19 9.7 13.3 0.0 77.0|UUUUssssss > | │
│ 20 6.1 12.2 0.0 81.6|UUUssssss > | │
│ 21 8.7 13.3 0.0 78.1|UUUUssssss > | │
│ 22 7.9 10.9 0.0 81.2|UUUsssss > | │
│ 23 6.6 11.6 0.0 81.8|UUUsssss > | │
│ 24 9.6 11.7 0.0 78.7|UUUUsssss > | │
│---------------------------+-------------------------------------------------+ │
│Avg 11.5 9.1 0.0 79.4|UUUUUssss > | │
│---------------------------+-------------------------------------------------+


So the NDB IO thread seems to be CPU bound.

...
ODirect=1
MinDiskWriteSpeed=30M
MaxDiskWriteSpeed=200M
MaxDiskWriteSpeedOtherNodeRestart=300M
MaxDiskWriteSpeedOwnRestart=500M
DiskIoThreadPool=8
FragmentLogFileSize=256M
InitFragmentLogFiles=SPARSE
NoOfFragmentLogFiles=300
RedoBuffer=128M
CompressedLCP=1
MaxBufferedEpochs=500 # Default 100
TimeBetweenLocalCheckpoints=20
TimeBetweenGlobalCheckpoints=1000
TimeBetweenEpochs=100
...

Is there anything we could do to improve this situation?