On Jun 1, 2020, at 5:10 PM, Mark Nelson <mnelson@redhat.com> wrote:

Hi Yiming,


Are you changing the overall data set size when you change the image size?  IE in your 40GB image test, is your data set 40x larger than in your 1GB image test?  
I’m using the same workload:
rw=randwrite
bs=4096
time_based=1
runtime=300
direct = 1
iodepth=48
Both run has the same run time 300s.

That would have various effects, including changing the number of onodes in the cache and the potential for cache misses hitting rocksdb and eventually the disk.  Having said that, with the default 4GB memory target I wouldn't expect you to have cache misses with typical RBD workloads even with a 40GB dataset on a single OSD unless you've tweaked the object size to be smaller or caused additional metadata per object in some way (EC, etc).


Theoretically you might be able to use lttng or jaeger tracepoints to track latency, or possible look at the perf counters.  Otherwise you might also be able to see something through wallclock profiling.
I tried the gdb wallclock profiling. The I can only see the fio and osd related time, not include the bluestore resources. Details please see here.
I added bunch of perf counters in BlueStore to track the latencies. I don’t see any suspicious counters. For locking behavior, is there any possible reasons for that? Really appreciated if you could point me which lock you mean in kv_sync_thread.

Thanks,
Yiming  
 I would probably look carefully at things happening in the kv sync thread since this is a random write workload and that's where I'd expect to see blocking behavior that could cause latency spikes like this.


Mark

On 6/1/20 1:50 PM, Yiming Zhang wrote:
Hi All,

I have noticed that different RBD image size can shape the bluestore latency differently. Is there baseline or guidance for choosing the image size?
Left: RBD image size is 1GB
middle: RBD image size is 40GB
Right: RBD image size is 1GB, RocksDB write buffer 10X default

4K randwrite on SSD with FIO. SSD is preconditioned and image is prefilled(20mins).
Red dot is L1 compaction and green dot is L0 compaction.


Let’s focus on the left graph. The smaller spikes are caused by compactions. The higher spikes seems to be caused by the BlueStore itself.
I suspect this could be related to RBD image size in someway.

Does anyone know what could the cause of the higher spikes? And how to debug it?

Also, what is the proper RBD image size for my test?

Please advice.

Thanks,
Yiming

_______________________________________________
Dev mailing list -- dev@ceph.io
To unsubscribe send an email to dev-leave@ceph.io
_______________________________________________
Dev mailing list -- dev@ceph.io
To unsubscribe send an email to dev-leave@ceph.io