FS of choice for max random iops ( Maildir )

freebsd at top-consulting.net freebsd at top-consulting.net
Fri Sep 16 21:39:59 UTC 2011

Quoting Terje Elde <terje at elde.net>:

> On 16. sep. 2011, at 16:18, freebsd at top-consulting.net wrote:
>>  Got a measly 74MB/sec.
> You can't ask for advice, get it, do something completely different,  
> and then complain that it didn't work.
> Neither can you ask people to donate their time, if you won't spend yours.
> In other words: if you won't listen, there's no point in us talking.
> However:
> Don't disable ZIL. Just don't. It's not the way to go. If you want  
> to know why, google will help.
> Also, you're making some assumptions, such as the ZIL being bad for  
> performance. That's not always the case. ZIL-writes are a rather  
> nice load for spinning metal storage. Even if you write through  
> cache, that can give you a boost on your real world workload.
> Which brings us to the third bit. You're benchmarking, not trying  
> real world loads. That's the load you'll have to worry about, and  
> it's the load zfs shines at.
> Thanks to the ZIL (the thing you're trying to kill, remember?) you  
> can convert seek heavy writes to sequential zil-writes, freeing up  
> disk bandwith for concurrent reads.
> If you want to test before spending money, try what Svein said. Set  
> up a small logical volume (preferrably smaller than your controller  
> cache, if it's large enough), then try that as a dedicated zil-device.
> Never tried that, but worth a shot.
> Terje

It's not about spending money or not. I really want to use ZFS for  
some of its features ( journaled, snapshots, etc ) but it has to be a  
good fit for me. I'm not ignoring the advice I am given, just taking  
it with a grain of salt disabling the ZIL is recommended - sometimes -  
for NFS.

As per hundreds of messages I've read from the Archive along with this  
page, http://wiki.freebsd.org/ZFSTuningGuide, it does appear that  
disabling the ZIL is  a solution for NFS. Yes, they still recommend  
SSD drives and I fully understand that. My point was the following:

Why is a sequential write test like dd slower on ZFS than on UFS ? The  
writes is already serialized so enabling/disabling the ZIL should have  
very little impact - which is indeed the case.

I even went as far as disabling the cache flush option of ZFS through  
this variable: vfs.zfs.cache_flush_disable: 1, since I already have  
the write cache of the controller. I've also set some other variables  
as per the Tuning guide but according to several benchmarks ( iozone,  
bonnie++, dd ) ZFS still comes in slower than UFS at pretty much  

Either I am missing something or there is something wrong with my setup.

More information about the freebsd-questions mailing list