[Gluster-users] increase write performance in gfs 3.5.2

[Rafi K C]

On Mon, 2015-02-09 at 13:58 +0100, Peter P. wrote:
> Thank you. And what are the other two jobs ?
2) Collapse small writes.
    if the write request is smaller, then it will collapse maximum up to
page size (128KB) .

3) Aggregate write requests
	Aggregate one or more write request and send as a vectored write. 

> 
> I have set the write-behind-window-size to 500 Mb and wrote a 100 Mb
> file, but the speed is still the same.
> Can you explain why ?

Once we lied to application, then we will start processing the lied
list. So the entries will be deleted if one request is fulfilled. Though
it is a 100Mb write there is no need of the same window-size. Default
window-size is 1 MB. You can see the performance enhancement if you
turned-off the write-behind translator. 

-Rafi KC 


> 
> 
> Regards
> 
> 2015-02-09 13:20 GMT+01:00 Peter P. <peter.prusi at googlemail.com>:
>         Thank you. And what are the other two jobs ?
>         
>         I have set the write-behind-window-size to 500 Mb and wrote a
>         100 Mb file, but the speed is still the same.
>         Can you explain why ?
>         
>         
>         Regards
>         
>         Peter
>         
>         
>         2015-02-08 12:04 GMT+01:00 Rafi K C
>         <muhammedrafi13 at gmail.com>:
>                 On Sat, 2015-02-07 at 20:41 +0100, Peter Prusi wrote:
>                 > Hello,
>                 >
>                 > I'm new to glusterfs and I dont understand 'the
>                 > performance.write-behind-window-size' option. I'm
>                 using glusterfs 3.5.2
>                 > and setup a volume over two nodes connected over 1
>                 GbE.
>                 > Actually the write performance is about 58 Mb/s.
>                 This means glusterfs
>                 > writes the data simultaneous on both nodes. I'm
>                 trying to increase the
>                 > write speed and I thought this option would cache
>                 the data
>                 > on one node and then distribute to the other. So i
>                 could double the
>                 > speed. I played around with this option, but it has
>                 no effect to the
>                 > write speed.
>                 
>                 Write-behind does three jobs with respect to a
>                 write-request. One among
>                 them is,
>                 
>                 it will lie to application, which issued the
>                 write-request about the
>                 completion of write and then will perform this write
>                 operation in
>                 background. Basically, For a single inode, it will lie
>                 to the
>                 application until current size of the total lied
>                 request which are not
>                 fulfilled on nodes becomes equal to the window size.
>                 This size is
>                 configurable through
>                 performance.write-behind-window-size .
>                 
>                 
>                 > What am I doing wrong or
>                 > is there another tweak to increase the write
>                 > speed ?
>                 
>                 I'm not familiar with write performance tweaks. Some
>                 one more familiar
>                 with this can comment on this.
>                 
>                 
>                 Regards
>                 Rafi KC
>                 >
>                 > Thank you
>                 > _______________________________________________
>                 > Gluster-users mailing list
>                 > Gluster-users at gluster.org
>                 >
>                 http://www.gluster.org/mailman/listinfo/gluster-users
>                 
>                 
>         
>         
> 
> 
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users