[zfs-discuss] ZFS and Power Loss

Adam Boyhan via zfs-discuss zfs-discuss at list.zfsonlinux.org
Tue May 5 12:54:04 EDT 2015


I have two hardware configurations setup. 

1. HP MSA 2040 using the HP supplied SSD's with a HP DL380p gen9 front end. I think these are based on one of the sandisk models. They are write endurance drives. 

2. Super Micro build with 4TB mechanical and a 800GB Intel DC3500 

I can reproduce the issue on both of the setups in both a physical environment and virtual. In my opinion recreating the issue in the virtual environment takes the issue of the SSD's built in cache out of the picture because they aren't truly loosing power. I am just hard stopping the VM which is in theory a power failure at the OS level. 

If the SSD for the zil are not fast enough to keep up, shouldn't they in theory just slow the application down? 


From: "zfs-discuss" <zfs-discuss at list.zfsonlinux.org> 
To: "zfs-discuss" <zfs-discuss at list.zfsonlinux.org> 
Cc: "nick pitlosh" <nickpitlosh at gmail.com> 
Sent: Tuesday, May 5, 2015 12:44:16 PM 
Subject: Re: [zfs-discuss] ZFS and Power Loss 

whats the zil on? 

On Tue, May 5, 2015 at 12:43 PM, nick pitlosh < nickpitlosh at gmail.com > wrote: 



tru dat. 

On Tue, May 5, 2015 at 11:48 AM, Gordan Bobic via zfs-discuss < zfs-discuss at list.zfsonlinux.org > wrote: 

BQ_BEGIN

I think the original point was that your SSD may not be as reliably in the power loss scenario as you think. Try disabling write caching on your ZIL SSD, set sync=always and test the application with O_SYNC. It is plausible that disabling write cache on the ZIL SSD may make the problem go away. 


On Tue, May 5, 2015 at 4:44 PM, Dan Swartzendruber via zfs-discuss < zfs-discuss at list.zfsonlinux.org > wrote: 

BQ_BEGIN
On 2015-05-05 11:40, Uncle Stoat via zfs-discuss wrote: 

BQ_BEGIN
On 05/05/15 16:04, Adam Boyhan wrote: 

BQ_BEGIN
That was what I was thinking. However, I still run into issues even 
when running sync=always. 



If you're losing data, then the problem is likely to be outside of ZFS, 
or your SSD isn't fast enough. 

BQ_END

Maybe I am confused, but this doesn't sound right to me. I thought if you were O_SYNC, the contract was that the kernel would not return until the write had committed to stable storage. If using an SLOG, the client is told the write completed even though it's only in SLOG still. If you power fail while that happens, on reboot, when the pool is imported, anything in the SLOG should be getting replayed and committed to spinning rust. Or am I missing something? 

_______________________________________________ 
zfs-discuss mailing list 
zfs-discuss at list.zfsonlinux.org 
http://list.zfsonlinux.org/mailman/listinfo/zfs-discuss 

BQ_END



_______________________________________________ 
zfs-discuss mailing list 
zfs-discuss at list.zfsonlinux.org 
http://list.zfsonlinux.org/mailman/listinfo/zfs-discuss 


BQ_END



BQ_END



_______________________________________________ 
zfs-discuss mailing list 
zfs-discuss at list.zfsonlinux.org 
http://list.zfsonlinux.org/mailman/listinfo/zfs-discuss 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.zfsonlinux.org/pipermail/zfs-discuss/attachments/20150505/0346560b/attachment.html>
-------------- next part --------------
_______________________________________________
zfs-discuss mailing list
zfs-discuss at list.zfsonlinux.org
http://list.zfsonlinux.org/mailman/listinfo/zfs-discuss


More information about the zfs-discuss mailing list