high disk %busy, while almost nothing happens

[Adam Vande More]

On Thu, Nov 26, 2015 at 10:46 PM, Eugene M. Zheganin <emz at norma.perm.ru>
wrote:

> Hi.
>
> On 26.11.2015 14:19, Eugene M. Zheganin wrote:
> > Hi.
> >
> > I'm using FreeBSD 10.1-STABLE as an application server, last week I've
> > noticed that disks are always busy while gstat shows that the activity
> > measured in iops/reads/writes is low, form my point of view:
> >
> >
> >   L(q)  ops/s    r/s   kBps   ms/r    w/s   kBps   ms/w   %busy Name
> >      8     56     50     520  160.6      6    286  157.4  100.2
> gpt/zfsroot0
> >      8     56     51   1474  162.8      5    228  174.4   99.9
> gpt/zfsroot1
> >
> > These %busy numbers arent't changing much, and from my point of view
> > both disks do very little.
> >
> The thing is, it was the compression. As soon as I cleared the gzip
> compression from busy datasets, %busy went down, almost to zero.
> Affected datasets were filled with poorly compressionable files, mostly
> archives or zlib-compressed data.
>

Data which isn't very compressible isn't a very great on a transparently
compressed filesystem.  Gzip is particularly bad at this.  LZ4 may have had
only a slight impact.  Setting gzip-1 would have also been less overhead
than the default gzip which I believe is gzip-6.


> And this is kind of counter-intuitive: one could think that worse-case
> scenario would be redundant CPU load, with constand disk i/o. In
> practice, otherwise, high disk %busy happens.
>

Well that's basically what you had.  And %busy is not really meaningful.
L(q) and ops are the ones to keep an eye on.

-- 
Adam