We have an A400 which has a few SVMs, one of which is used for a specific CIFS workload that is showing an avg. latency of over 20ms.
We have the havester running, and from that we can see that the workload is about 800 write iops and 300 read iops. And the main CIFS operation types are open/close (20K/sec) and GetAttr (15K/sec)
There are about 4.000 open files.
The read and write latency isn't that bad (under 2ms) but I guess the CIFS operations is what is triggering the high avg. latency?
Anyway, the guys responsible have been complaining about the higher latency. Upgrading from 9.12 to 9.14 actually made the latency a but worse...
We have disabled all QoS on the cluster, and we are considering adding QoS to this specific workload, but I have not had great experience with QoS, so I would rather investigate if it was possible to optimize the CIFS SVM in some way? So if anyone have some suggestions as to what to optimize. The current CIFS role is basic and has not been tampered with.
#CIFS performance optimization
tall fossil
weary sable
is there any impact releated to latency? Otherwise the problem with low activity on volumes is that the avg. latency is not really ueful: https://kb.netapp.com/onprem/ontap/Performance/Why_is_a_workloads_latency_high_when_the_IOPS_are_low
tall fossil
Well they changed the backend workloads so they no longer have so many files open at one time... this seems to have lowered the avg. latency quite a bit. From what I can understand from the KB, this is related to workloads with low activity which is of cause all relative, but this workload is writing and reading all the time because it is an ELK Stack that is analyzing logs from many different systems... for now it is not that critical any longer, but we will keep an eye on it because the load will become more than double... 🙂
undone niche
tall fossil
The 20ms is the average latency as shown in the SVM -> CIFS dashboard in NABox/Havester. The read and write latency is always lower than the average, I guess this is because of the many setattr, getattr etc. operations?
azure sundial
the real question: are clients complaining?