Hi Nathan, One of the things to check is whether Solaris thinks the device is volatile and therefore needs a cache flush for each write. Unless it's one of the relatively few devices that are hardcoded into the relevant tables in sd driver, Solaris by default will treat it as volatile. You could check the un_f_suppress_cache_flush bit for the device, or just go ahead and put an appropriate entry into sd.conf. For example (for a different device):
sd-config-list = "ATA MARVELL SD88SA02","emulation-rmw:0, cache-nonvolatile:true"; "disksort:false" might also be appropriate I'd also suggest for testing purposes to "precondition" the device by writing to it sequentially several times using maybe 1M iosize. [Working on the assumption you don't have a convenient means of redoing a low-level format.] Sustained write performance is heavily dependent on there being empty erase blocks available, which a series of sequential writes should help to achieve. Might not be relevant, but some devices lie about their internal geometry, claiming to be 512B when they're really 4K. Richard. On 12/01/2013 05:56 PM, Nathan Kroenert wrote: > Hi folks, > > In making an attempt to be ready for this month's meeting (which might > need to change date, as my son's Christmas Carols is on the 18th), I > have been playing with some different ZIL devices. > > Thus far, I'm not nearly as happy as I was last time I played with this > - and whilst I'm betting that it's my actual system that's the problem, > I thought I'd throw it out there to see if anyone had any thoughts > before I *really* start sinking time into this... > > Bottom line is that I'm getting crappy woeful performance from the SSD. > Sure, it's substantially better in terms of random IOPS than an HDD, but > it's nowhere where what I had previously come to expect. > > dd'ing sequentially to /dev/rdsk/cxtxdxs0 with 4K blocks gives me only > about 1000 IOPS, which is about 90x less than the quoted maximum - and > is hardly setting the world on fire. > > The SSD is a Samsung 840 pro - and I was expecting it to be better than > my trusty old Intel 520 series jobbie, which is presently anywhere up to > 5X faster than the sammy, which is, according to the box, supposed to be > able to do 90,000 random write IOPS and about 500 odd MB/s. It's plugged > into a Gigabyte motherboard, with Sata3 controllers - and is training up > at Sata 3 signalling speed. > > Things to know: > - Maxphys: > > maxphys::print -td > int 0t131072 > > - emulation-rmw isn't set in sd.conf (I'll check it out at some point > - and explicitly set it, but it shouldn't be getting in my way if I'm > doing 4K aligned block writes.) > > - Slice 0, which is what I'm using for my testing is indeed 4K block > aligned: > Part Tag Flag First Sector Size Last Sector > 0 usr wm 256 223.56GB 468845710 > So - 256 * 512 bytes == 131072 ----> > 131072 / 4096 == 32 ----> So we are 4K aligned. > > - When using the samsung, I seem to get *fewer* IOPS when I dd > directly to the slice than ZFS gets when using it as a ZIL. I really > don't understand this one... yet. ;) (The Intel one gives me more IOPS > using dd than ZIL, which more like what I expected.) > > - I'm not seeing a bazillion errors in an iostat -En. I am seeing some > illegal requests - but a reasonably low number and they aren't > skyrocketing when I blast the drives. (I really should look into why I'm > getting illegal requests to the drives though - I'm seeing some issues > on all my drives... More than I'd like) > > - I have tried jacking with things to get the CPU's to spend more time > in their full power c-state. It makes about a 10% difference, so not the > multi-thousand % increase I'm looking for. ;) > > Should anyone have any thoughts, I'm open to them. > > I'll be poking more - but thought it would be worth an ask. > > Cheers, > > Nathan. > > > > _______________________________________________ > msosug mailing list > [email protected] > http://mexico.purplecow.org/m/listinfo/msosug > -- ============================================================================ ,-_|\ Richard Smith PAE / \ Oracle Phone : +61 3 8616 3300 [email protected] Direct : +61 404 815 885 \_,-._/ 417 St Kilda Road Fax : +61 3 8616 4500 v Melbourne Vic 3004 Australia =========================================================================== _______________________________________________ msosug mailing list [email protected] http://mexico.purplecow.org/m/listinfo/msosug
