[sheepdog-users] Concern about sheepdog performance

[Liu Yuan]

On 12/18/2012 06:07 PM, Valerio Pachera wrote:
> Hi, I found this page of the archive
>   http://www.mail-archive.com/sheepdog@lists.wpkg.org/msg00932.html
> 
> There's reported about my same data (5-10M/s).
> It also reports that disabling barrier when mounting filesystem, we
> should gain lot's of performance.
> I wanted to try that!
> 
> I updated my sheepdog to latest develop version (0.5.5_19_g395c05c).
> 3 nodes, copies 2.
> I also tried to add allocsize=4M because I know sheepdog uses file of 4M.
> Is it a good idea?
> 
> mount (xfs): -o noatime,barrier=0,allocsize=4M
> sheep -D /mnt/sheepdog
> kvm -drive file=sheepdog:test,cache=none
>   16M/s
> 
> (the same, but without spcifing cache on qemu)
> mount (xfs): -o noatime,barrier=0,allocsize=4M
> sheep -D /mnt/sheepdog
> kvm -drive file=sheepdog:test
>   15.3M
> 

In both cases, you don't use cache at all, so 'cache=type' is meanless
actually they are the same thing to sheepdog.

As I mentioned before, use 'virtio' interface would result in better
performance and object cache (writeback mode) + journaling will give you
the best combination for performance.


> ---------
> 
> Now I wanted to run sheep without '-D'; I shutdown the cluster, re-run
> 'sheep /mnt/sheepdog' but only one node started the daemon.
> On the other two, in sheep.log we find:
> 
> Dec 18 11:02:57 [main] send_join_request(1014) IPv4 ip:192.168.2.41 port:7000
> Dec 18 11:02:57 [main] farm_init(118) xattrs are not enabled on
> /mnt/sheepdog/obj/

xattrs isn't enabled on the device so sheep exit.

> Dec 18 11:02:57 [main] crash_handler(324) sheep pid 15494 exited unexpectedly.
> Dec 18 11:02:57 [main] jrnl_recover(230) opening the directory
> /mnt/sheepdog/journal/
> Dec 18 11:02:57 [main] jrnl_recover(235) starting journal recovery
> Dec 18 11:02:57 [main] jrnl_recover(291) journal recovery complete
> 
> Dec 18 10:54:28 [main] send_join_request(1014) IPv4 ip:192.168.2.43 port:7000
> Dec 18 10:54:28 [main] farm_init(118) xattrs are not enabled on
> /mnt/sheepdog/obj/
> Dec 18 10:54:28 [main] crash_handler(324) sheep pid 13656 exited unexpectedly.
> Dec 18 10:54:28 [main] jrnl_recover(230) opening the directory
> /mnt/sheepdog/journal/
> Dec 18 10:54:28 [main] jrnl_recover(235) starting journal recovery
> Dec 18 10:54:28 [main] jrnl_recover(291) journal recovery complete
> 
> *On the node with the running daemon*
> Dec 18 10:51:37 [main] send_join_request(1014) IPv4 ip:192.168.2.42 port:7000
> Dec 18 10:51:37 [main] main(616) sheepdog daemon (version
> 0.5.5_19_g395c05c) started
> Dec 18 10:51:37 [main] update_cluster_info(799) status = 2, epoch = 0,
> finished: 0
> Dec 18 10:51:41 [main] sd_check_join_cb(974) 192.168.2.43:7000: ret =
> 0x0, cluster_status = 0x2
> Dec 18 10:54:03 [main] sd_check_join_cb(974) 192.168.2.43:7000: ret =
> 0x0, cluster_status = 0x2
> Dec 18 10:54:25 [main] sd_check_join_cb(974) 192.168.2.43:7000: ret =
> 0x0, cluster_status = 0x2
> 
> Is it better to not use the development verion for testing?
> 

Either is okay. But I'd suggest use master for test because in most of
the time, the master is more stable as time goes by.

Thanks,
Yuan