Complexity of toggling write-behind.

Complexity of toggling write-behind.

[greg]

Hi Neil/linux-raid, hope the week is going well for everyone.

We implement a lot of geographically mirrored storage using Linux
RAID1 on the 'front-side' to access Linux servers running SCST for
storage backends.  One of the issues we are addressing are methods to
reduce I/O latency from the initiator's perspective when various
housekeeping operations are being done on the Linux backends, ie. RAID
check operations/re-synchronizations.

Write-mostly support is very helpful for reducing any read latencies.
To fully complete the loop it would be useful to leverage asynchronous
write-behind to eliminate write latency effects.  We use write-intent
bitmaps on the initiators so we are already preparted for that.

It is straight forward to toggle writemostly via the
/sys/mdN/md/rN/state pseudofiles but a review of the MD driver code in
at least 2.6.32.16 would suggest that write-behind is not readily
togglable.  That may have changed in most recent kernels but I don't
remember seeing any changelogs with your 'git pull' requests which
would suggest that.

Is this a function of support being a can of worms or just lack of
cycles?  If it is the latter I would be willing to take some
directional suggestions and see if we could get a prototype
implementation up and running.

Thanks much for any insight you may have.

Best wishes for a productive remainder of the week.

As always,
Dr. G.W. Wettstein, Ph.D.   IDfusion.org
4206 N. 19th Ave.           Unified health identity architecture.
Fargo, ND  58102
PH: 701-281-1686
FAX: 701-281-3949           EMAIL: greg@idfusion.org
------------------------------------------------------------------------------
"We don't know the OS that God uses, but the Vatican uses Linux."
                                -- Sister Judith Zoebelein
                                   Vatican Webmaster

Re: Complexity of toggling write-behind.

[Neil Brown]

On Wed, 13 Oct 2010 13:48:50 -0500
greg@enjellic.com wrote:

> Hi Neil/linux-raid, hope the week is going well for everyone.
> 
> We implement a lot of geographically mirrored storage using Linux
> RAID1 on the 'front-side' to access Linux servers running SCST for
> storage backends.  One of the issues we are addressing are methods to
> reduce I/O latency from the initiator's perspective when various
> housekeeping operations are being done on the Linux backends, ie. RAID
> check operations/re-synchronizations.
> 
> Write-mostly support is very helpful for reducing any read latencies.
> To fully complete the loop it would be useful to leverage asynchronous
> write-behind to eliminate write latency effects.  We use write-intent
> bitmaps on the initiators so we are already preparted for that.
> 
> It is straight forward to toggle writemostly via the
> /sys/mdN/md/rN/state pseudofiles but a review of the MD driver code in
> at least 2.6.32.16 would suggest that write-behind is not readily
> togglable.  That may have changed in most recent kernels but I don't
> remember seeing any changelogs with your 'git pull' requests which
> would suggest that.
> 
> Is this a function of support being a can of worms or just lack of
> cycles?  If it is the latter I would be willing to take some
> directional suggestions and see if we could get a prototype
> implementation up and running.


In 2.6.33 and later you can write to /sys/block/mdX/md/bitmap/backlog

When there is a bitmap and backlog is non-zero, writes to read-mostly devices
a 'behind' writes (up to the limit of 'backlog').

NeilBrown


> 
> Thanks much for any insight you may have.
> 
> Best wishes for a productive remainder of the week.
> 
> As always,
> Dr. G.W. Wettstein, Ph.D.   IDfusion.org
> 4206 N. 19th Ave.           Unified health identity architecture.
> Fargo, ND  58102
> PH: 701-281-1686
> FAX: 701-281-3949           EMAIL: greg@idfusion.org
> ------------------------------------------------------------------------------
> "We don't know the OS that God uses, but the Vatican uses Linux."
>                                 -- Sister Judith Zoebelein
>                                    Vatican Webmaster