[OmniOS-discuss] slog limits write speed more than it should

[Richard Elling]

On Nov 14, 2014, at 3:06 PM, Rune Tipsmark <rt at steait.net> wrote:

> I get only about half the bandwidth with Sync=Always compared to Sync=Disabled.
> Using an SLC device that should perform better, its rated as 750 MB/sec, I only get something like 60% of that at best of times. If there is a way to stripe a ZIL it would be great. I have enough to mirror two stripes and get the speed I want.

Technically, you can't mirror stripes in ZFS. You can stripe mirrors. For write workloads (eg slog)
a mirror is always at least slightly slower than unmirrored disks. For read workloads, mirrors are
demonstrably faster than unmirrored disks.
 -- richard

>  
> Copying ~60GB from one LUN to another on same ZFS box.
>  
> Sync=Always:
> <image001.png>
>  
> Sync=Disabled
> <image002.png>
>  
> -----Original Message-----
> From: Rune Tipsmark 
> Sent: Friday, November 14, 2014 11:53 AM
> To: Rune Tipsmark; Bob Friesenhahn
> Cc: omnios-discuss at lists.omniti.com
> Subject: RE: [OmniOS-discuss] slog limits write speed more than it should
>  
> Is there a way to stripe two block devices and use them as log?
> I tried following this one https://blogs.oracle.com/bilke/entry/raid_0_stripe_on_solaris
> But I cannot use the device in ZFS - getting the following error: cannot use '/dev/md/rdsk/zil1d0': must be a block device or regular file
>  
> Br,
> Rune
>  
> -----Original Message-----
> From: OmniOS-discuss [mailto:omnios-discuss-bounces at lists.omniti.com] On Behalf Of Rune Tipsmark
> Sent: Friday, November 14, 2014 9:47 AM
> To: Bob Friesenhahn
> Cc: omnios-discuss at lists.omniti.com
> Subject: Re: [OmniOS-discuss] slog limits write speed more than it should
>  
> Well that sucks... I guess one more reason to move to NV-Dimms to replace slow SLC cards.
> Br,
> Rune
>  
> -----Original Message-----
> From: Bob Friesenhahn [mailto:bfriesen at simple.dallas.tx.us]
> Sent: Friday, November 14, 2014 6:48 AM
> To: Rune Tipsmark
> Cc: omnios-discuss at lists.omniti.com
> Subject: Re: [OmniOS-discuss] slog limits write speed more than it should
>  
> On Fri, 14 Nov 2014, Rune Tipsmark wrote:
>  
> >
> > Looking at the first link he clearly states that ZIL is written to in
> > a round robin fashion, I was under the impression that 2 log devices would then be faster than 1…unless mirrored of course.
>  
> This could be a mistaken impression.  The slog commits each write before proceeding to the next write.  If the involved SSDs have a fixed minimum write latency, then this would limit the maximum transaction rate regardless of the number of SSDs involved.  There could be an advantage to more SSDs if the additional time between writes allows the SSD to more effectively prepare for the next write and reduce the write latency.
>  
> Bob
> --
> Bob Friesenhahn
> bfriesen at simple.dallas.tx.us, http://www.simplesystems.org/users/bfriesen/
> GraphicsMagick Maintainer,    http://www.GraphicsMagick.org/
> _______________________________________________
> OmniOS-discuss mailing list
> OmniOS-discuss at lists.omniti.com
> http://lists.omniti.com/mailman/listinfo/omnios-discuss
> _______________________________________________
> OmniOS-discuss mailing list
> OmniOS-discuss at lists.omniti.com
> http://lists.omniti.com/mailman/listinfo/omnios-discuss

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://omniosce.org/ml-archive/attachments/20141114/803689ff/attachment-0001.html>