zfs very poor performance compared to ufs due to lack of cache?
avg at freebsd.org
Wed Sep 15 15:15:56 UTC 2010
on 15/09/2010 18:04 Steven Hartland said the following:
> ----- Original Message ----- From: "Andriy Gapon" <avg at freebsd.org>
>>> Indeed. Where would this need to be addressed as ufs doesn't suffer from this?
>> In ZFS. But I don't think that this is going to happen any time soon if at all.
>> Authors of ZFS specifically chose to use a dedicated cache, which is ARC.
>> Talk to them, or don't use ZFS, or get used to it.
>> ARC has a price, but it supposedly has benefits too.
>> Changing ZFS to use buffer cache is a lot of work and effectively means not using
>> ARC, IMO.
> Hmm, so taking a different track on the issue is the a way to make sendfile use data
> directly from ARC instead of having to copy it first?
Well, theoretically everything is possible, but I am not sure if it's feasible.
It's a lot of work anyways, it should be a very specialized sendfile and a lot if
inter-layer knowledge and dependencies.
Don't hold your breath for it.
>> Well, I thought that you hurried when you applied the patches and changed the
>> settings at the same time. This made it impossible for you to judge properly what
>> patches do and don't do for you.
> No hurry just applying the patches that where suggested, retest, apply new retest
> etc but
> in parrallel been reading up on the arc tunables.
>>> Now we have a very simple setup so we can make sensible values for min / max but
>>> it still means that for every file being sent when sendfile is enabled:
>>> 1. There are two copies in memory which is still going to mean that only half the
>>> amount files can be successfully cached and served without resorting to disk IO.
>> Can't really say, depends on the size of the files.
>> Though, it's approximately a half of what could have fit in memory with e.g.
>> UFS, yes.
> Out of interest if a copy of the data is being made from ARC whats ties those
> two copies together, in order to prevent the next request for the same file having to
> create a third copy etc...
Read about FreeBSD VM, particularly about a notion of VM object.
>>> 2. sendfile isn't achieving what it states it should be i.e. a zero-copy. Does
>>> this explain
>>> the other odd behaviour we noticed, high CPU usage from nginx?
>> sendfile should achieve zero copy with all the patches applied once both copies of
>> data are settled in memory. If you have insufficient memory to hold the workset,
>> then that's a different issue of moving competing data in and out of memory. And
>> that may explain the CPU load, but it's just a speculation.
> Yes, more investigation needed ;-)
>> At present I don't see any other way but brute force - throw even more RAM at the
>> Perhaps, a miracle would happen and someone would post patches that radically
>> change ZFS behavior with respect to caches. But I don't expect it
> Or alternatively make sendfile work directly from ARC, would that be possible?
I'll be glad to review the patches :-)
> Thanks for all the info :)
You are welcome.
You did a lot of testing and investigative work here and I hope this thread will
be useful for other people researching the topic.
More information about the freebsd-fs