[OmniOS-discuss] Pliant/Sandisk SSD ZIL

[Jim Klimov]

On 2014-02-23 06:14, Richard Elling wrote:
> On Feb 18, 2014, at 10:40 AM, Jim Klimov <jimklimov at cos.ru> wrote:
>
>> On 2014-02-18 08:05, Richard Elling wrote:
>>> Fortunately, most real workloads are not tar -x.
>>
>> Oh, alas, on build farms they are. And the hordes of files produced
>> as a result of "make" are not far behind. Oh, well, often they are
>> behind - and subsequent accesses fail for a few seconds, just until
>> something syncs down the pipe and back to the client. So the typical
>> mantra is "gmake -j -k all; gmake all" for producing as many objects
>> as we can first, and then cleaning up the mess sequentially :)
>
> For build farms, why do you care about NFS sync writes? This is the
> canonical case for sync=disabled, because there is no unique data.
>   — richard

Interesting suggestion :)

In one example I have, the builds from several nodes are done in
subdirectories of a user's home directory, as well as edition of
some files (sources, recipes/makefiles) in the process. And being
a home, it is sync'ed per defaults.

But I might investigate if dedicating a sub-dataset without sync
would help and won't cause races (i.e. edit a file on one host,
initiate a build quickly on another, and compile the obsolete
version which was cached by NFS client), so thanks :)

I hope, NFS cached-data syncs and locks, and ZFS write-syncs are
not very related in this case (i.e. zfs sync=disabled does not
influence co-ordination of NFS data between hosts), right?

Thanks,
//Jim