Since zfs uses a private cache (the ARC) for most of its file caching,
rather than VM pages, the "caching" or not doesn't make that much
difference. On the other hand, ufs does use VM pages, so the caching
options makes quite a bit of difference.
The fs_flush script is used to flush the ZFS file cache by exporting
and then importing the pool. In all the released versions this is
independent of whether the caching attribute is enabled or not. (The
about-to-be-released v1.4.7 will not run fs_flush if all filesets have
"cached" set).
Note that for fs_flush to work you have to be logged in as root, and
your data directory (where the filesets are created) has to be on a
non root pool (an issue if you are using zfs as the root filesystem).
I believe that Solaris always converts reads and writes into file
mapped I/O, so if the cached attribute is missing or set to "false"
the VM pages involved in that will be flushed, so you will see a small
difference in performance, but because the files still live in the
ARC, it is much smaller than with UFS, where the files have been
completely flushed from memory.
Drew
On Apr 28, 2009, at 12:20 PM, Dearl Neal wrote:
Anyone have a clue as to why the results from
large_db_oltp_Nk_cached and large_db _oltp_Nk_unchached workloads
would be about the same for a zfs filesystem. .. Again, the system I
am testing with is a V240, 2...@1503mhz, 4GB RAM, using MPxIO, a single
lun presented from from IBM SVC with EMC storage on the back end,
and 5.10 u6, filebench 1.4.4 .
The workloads for ufs show the uncached op/s are 2xcached op/s...
similiar results for the bandwidth (MB/s) .
Just another interesting note with my testing.. 1) zfs biglun with
SMI label is performing 2x faster that with an EFI label..2) with
zfs recordsize=128k and testing with a large_db_oltp_128k_cached
workload for filebench.. I am seeing ~1700 op/s and 200MB/s
throughput! .. anyway, I am seeing some pretty interesting results.
Thanks in advance.
-D
Message was edited by: dneal
--
This message posted from opensolaris.org
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
