Long pauses during write

[lundman]

Certainly interesting findings. Really one would expect it to be fast while ARC is filling, then slow down as it is full - that is how traditional caching works. I guess we have a little rebel ARC going on at the moment. We are happier with master's ARC, but that doesn't mean we are entirely happy with it yet. Could be there is still some odd starvation going on during ARC filling.

Having said that, we are currently looking at huge changes done by rottegift, where we finally use OSX's UBC with arc, so everything is about to be different But with some nice sideffects.

[mkush]

I decided to do my tests with regular-old 1.7.0 also. Turns out the performance is no worse than the master code, in fact maybe a couple of seconds better. And the effect of the ARC is the same as well; when set to 1GB, performance is zippy-fast, when set to 10GB, it is dreadfully slow and seemed to rapidly improve once the copy reached the size of the ARC. So my findings do not support performance improvements in the master code (doesn't mean they aren't there for other use cases). And my findings also show that the odd ARC issue is not new to the master code. So my advice at this point to those running pure SSD pools (at least over Thunderbolt) is to set your ARC small, say 1GB. Smaller still may be possible but too small seemed to not help.

[lundman]

Thanks for reporting your findings. We should ramp up the performance improvement part of the project now that it is mostly stable

[mkush]

My pleasure. It's the least I can do to repay your hard work. If I can do any further testing with my setup (SSD pool over Thunderbolt) please let me know. You may need to PM me (is that even possible?) since I don't always frequent the boards here.

[e8vww]

Thanks for reporting your findings. We should ramp up the performance improvement part of the project now that it is mostly stable

I am getting really bad performance from my 3x 8tb mirrors. 15mb/s write speed over sata with brand new drives. It contantly spikes up to 60mb/s but the average over hours is 15mb/s. Any idea what this could be caused by? Test copying using large .mkv video files. Thanks for your help.

EDIT: Reducing the arc_max to 1gb really sped up writes for me and reduced the spiking.

I'm noticing that a copy between pools 1 and 2 will cause a download speeds from the internet to pool 3 to become erratic. The cpu is hardly being used. Is there a way to assign higher priority to zfs?

It seems like there is not enough processing power within the zfs process to handle all the data.

[mkush]

Just wanted to update my old thread. After some time off the project, I've put my array, described above, into production. MacBook Pro is running version 1.7.2. I didn't try setting the ARC value to something "correct," just left it at 1GB as before. Just copied a 215.6GB folder (iPhone photos and videos), using the Finder, from internal boot SSD to ZFS array in 360 seconds, meaning 600 MB/s. This is almost exactly the performance measured before (which means I'm happy about it). If I get some time I may set the ARC to a normal value and try it again to see if the problem still exists. If I don't post back within a few days it means that the problem is still there (writes drastically slowing down once ARC is full).