Hi Chip,

On May 27, 2014, at 7:31 AM, Schweiss, Chip via illumos-discuss 
<[email protected]> wrote:

> 
> On Tue, May 27, 2014 at 8:46 AM, Bob Friesenhahn 
> <[email protected]> wrote:
> On Tue, 27 May 2014, Schweiss, Chip wrote:
> 
> Actually, with ZFS more free pages, does improve throughput, depending on how 
> aggressive the SSDs garbage collector is.  
> Since ZFS will write to every available block on the disk, after filling 
> once, every write afterwards is overwriting a
> previously written logical block.   If the SSD has enough spare blocks and an 
> aggressive enough garbage collector, there
> 
> The above statement is contrary to what Richard Elling said about zfs 
> behavior.  Richard Elling said that zfs allocates from lower-addressed 
> sectors first and so a non-full pool is unlikely to ever use upper-addressed 
> sectors.
> 
> It is of course easy to test actual zfs behavior using a dtrace script.
> 
> This is getting a bit off topic but quite important as to the reason you 
> would want to under provision an SSD.
> 
> I have no doubt, Richard, knows the internals of ZFS better than me.   I have 
> however done several long running IOzone tests against Samsung 840 Pro SSDs 
> sliced at 10%  intervals between 60% and 100%.   IOzone only allocated a 
> small percentage of the total.   It was always clear at 90% and 100% when the 
> total pool capacity had been written as performance would drop off, no matter 
> what the write size.   Between each run, all the SSDs were secure erased, 
> re-partitioned and allowed to sit idle for 1 hour so that all blocks were 
> zero when starting.   

Interesting results. Were you testing the "disk" or "disk with ZFS"?

> 
> Below 80% and only small write size tests would outrun the garbage collector. 
>   Even at 60% 4k random writes overran the the garbage collector and 
> performance tanked.   I never let a 4k test run to completion because they 
> ran for over 6 hours when larger block size tests would complete in less than 
> 1 hour.  
> 
> Any SSD pool, knowing the work load and characteristics of the SSD being used 
> is paramount to tuning it to the best performance.   Never expect the 
> manufactures published performance of any SSD for real workloads.  

SNIA's SSS-PTS is quite good at characterizing random block device performance.
Some vendors publish SSS-PTS results, encourage more to do so :-)
 -- richard

> 
> Back to the original question, slicing while not intuitive when first 
> encountered, does the job of keeping a reserved space from ZFS.   I used the 
> following format script to create an EFI partition and slice the SSD.  Adjust 
> to your particular SSD:
> 
> fdisk
> n
> 1
> 1
> 80
> y
> 6
> q
> 
> -Chip
> 
> 
> 
> illumos-discuss | Archives  | Modify Your Subscription         

--

[email protected]
+1-760-896-4422






-------------------------------------------
illumos-discuss
Archives: https://www.listbox.com/member/archive/182180/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4
Powered by Listbox: http://www.listbox.com

Reply via email to