I would expect to have the cache hit rate higher. My first guess would be your hitting cache so the arc isn't refilling.

If you want to get down and dirty, you can go into mdb and look at memstats to see what is in ram.

as root

mdb -k

:memstat

It will show what's using ram.

Hi Linda

The output is

Page Summary Pages MB %Tot ------------ ---------------- ---------------- ---- Kernel 6500143 25391 39% ZFS File Data 7991562 31217 48% Anon 132656 518 1% Exec and libs 1169 4 0% Page cache 4763 18 0% Free (cachelist) 8376 32 0% Free (freelist) 2130965 8324 13%

Total 16769634 65506 Physical 16769633 65506

Så almost a even split between kernel and ZFS File Data(is this ARC or does it reside in Kernel)?

so since you are already there..

we can look at what's going on in arc

inside nmc on cli

show performance arc

I ran it breafly and here is the output of that

Current ARC Size Min ARC Size(zfsarcmin) Max ARC Size(zfsarcmax) 19535MB 8060MB 64482MB

Cache hits and misses (total): Cache Hits: 77% Cache Misses: 16% Cache hits by type: Demand Data: 83% Prefetch Data: 6% Demand Metadata: 9% Prefetch Metadata: 0% Cache misses by type: Demand Data: 63% Prefetch Data: 36% Demand Metadata: 0% Prefetch Metadata: 0%

Can you show me another thing?

kstat -pn arcstats

It looks like evict might be high, meaning something is telling the arc to not fill... If so, i might have to file a bug.

Hi Linda

I've attached the output in an txt file for you.

Regards

Denis

Hi Linda

It's very aggressive in it's flushing the cache quite alot..

Is there a setting where we could tell it to be less aggressive?

Also our cache hit ratio is down to 80%

Regards

Denis

denis, this is a really interesting problem. i would love to have the support guys look at this, any chance you have a support contract?

if not don't worry..

The other odd thing about this is that the arc is showing up at 44gb. evict get larger over time.

Do you know how this started? What caused the flush? Did you have a large data transfer that created the memory pressure to flush?

Hi Linda

We do have a support contract. It is registered to deb(at)akait.dk

I dont know what caused it as all was fine when we went home for the weekend. However when we arrived at work on monday the cache was flushed and sitting at about 12gb. No work was conducted during the weekend.

Regards

Denis

Can you open a case? I would really like them to be able to take a closer look at this. It looks like a bug.