Replication speeds?

[Wesley Craig]

Obviously looking at more iostat information would give a better  
idea, but I'd estimate that you are NOT I/O bound.  Sorry I can't  
give you absolute numbers from UM, but I can share a patch that we  
wrote that we believe has increase sync throughput substantially, as  
evidenced by the lack of a sync backlog which we were getting before  
we added the patch.

The patch itself is pretty simple.  As with most TCP-based request- 
response protocols, the Cyrus sync protocol suffers from artificial  
"Nagle" delays.  See:

	http://en.wikipedia.org/wiki/Nagle's_algorithm

for what's good about Nagle's algorithm.  The patch disables Nagle  
delays in both sync_client and sync_server.

:wes

-------------- next part --------------
A non-text attachment was scrubbed...
Name: nodelay.diff
Type: application/octet-stream
Size: 1573 bytes
Desc: not available
Url : http://lists.andrew.cmu.edu/pipermail/info-cyrus/attachments/20070509/dedf6707/nodelay.obj
-------------- next part --------------

On 09 May 2007, at 11:05, Nik Conwell wrote:
> What sort of rates are you all getting for replication?
>
> At 2.3.7 for a manual sync_client for a user, I'm seeing ~35MB/ 
> minute across a 100M net to a linux SW RAID 1 pair of U320 disks.
>
> Is this speed typical or abysmal?
>
> Disks appear to be holding me back:
>
> Device:    rrqm/s wrqm/s   r/s   w/s  rsec/s  wsec/s    rkB/s     
> wkB/s avgrq-sz avgqu-sz   await  svctm  %util
> dm-0         0.00   0.00  0.00 429.96    0.00 3439.65     0.00   
> 1719.83     8.00     7.24   16.75   1.86  80.01
>
> Production (40,000+ users) would probably be FC and gigE...