cache hit % - from sysstat

[m_lubinski]

Hello,

I have question about this vaule from sysstat command - is it best to see this % as lowest as possible (~0) or rather the bigger percentage the better for your performance?

I do see Cache hit on level 75-90%, and I don't really know if what I see means, that it's good or bad...

If anyone could also say if CP time on level ~84-95% is good or bad (and if bad - how to further "tune up" filer, I will really appreciate it.

Thanks for any tips on that

[aborzenkov]

Cache hit – the bigger percentage, the better (100% means all data is served from RAM, which is the fastest component). But it is quite unrealistic. Low cache hit rate would mean mostly random access. I do not think 90% is that bad. 75% is a bit low but again – it depends on workload.

Consistent CP time of 95% means your filer barely has time to flush data on disks. You need to further analyze where data comes from (clients or internal activities); maybe you have some hot spots that slow down disk writes. The value itself is not bad but it does not leave much headroom for increasing load.

[m_lubinski]

ok, clear with cache hit, then this CP time, i guess this should be less than 50% ?

aborzenkov wrote:
You need to further analyze where data comes from (clients or internal activities); maybe you have some hot spots that slow down disk writes.

how to approach that to be completely sure where data comes from? and about second part (hotspots) - how to verify that?

thank you very much for your help.

[aborzenkov]

Compare incoming client data amount with disk write amount; this will give a hint whether all disk writes result from clients activity or are generated internally.

For host spots - you could use statit to check IO distribution across disks.

statit is advanced level command and so not really documented, nor is there any tool to automate analysis. The advantage is, it gives high level overview of filer activity in human readable form. The same data is available as various performance counters using stats command which may be easier to visualize.

[roman_verysell]

For 'Cache hit' a bigger value is beter, it means what data comes from cache, instead from disks in 95%'s (for example) of requests.