Re: What is rx_processing_limit sysctl for Intel igb NIC driver?
On Sep 4, 2012, at 7:13 PM, John Baldwin j...@freebsd.org wrote: On Sunday, September 02, 2012 10:41:15 pm Andy Young wrote: I am tuning our server that has an Intel 82576 gigabit NIC using the igb driver. I see a lot of posts on the net where people bump the rx_processing_limit sysctl from the default value of 100 to 4096. Can anyone tell me what this is intended to do? If you have multiple devices sharing an IRQ with igb (and thus are not using MSI or MSI-X), it forces the driver to more-or-less cooperatively schedule with the other interrupts on the same IRQ. However, since igb uses a fast interrupt handler and a task on a dedicated taskqueue in the non-MSI case now, I think it doesn't even do that. It should probably be set to -1 (meaning unlimited) in just about all cases now. -- John Baldwin ___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to freebsd-hardware-unsubscr...@freebsd.org And setting it to -1 gave a nice performance improvement in some tests that I did recently. AFAIR only after setting this to -1 I was able to reach 10gig speed using iperf on two directly connected machines with ix(4) 82599___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to freebsd-hardware-unsubscr...@freebsd.org
Re: Support for Fusion IO drives?
Fron the zpool man page: By default, the intent log is allocated from blocks within the main pool. However, it might be possible to get better performance using separate intent log devices such as NVRAM or a dedicated disk. I was also contemplating the idea of a fast SSD on PCIe as a ZIL and L2ARC. And given the fact that the SSD will not suffer from the different types and locations of IO requests, maybe it makes sense to go with a big SSD, and partition it for a small ZIL partition, and the rest for L2ARC. Anyone tried that? Regards, Nikolay On Aug 29, 2012, at 2:47 AM, Andrew Young ayo...@mosaicarchive.com wrote: Thanks for the great feedback Josh! The optimum size for an ssd zil device was still an open question for us. I'm really glad to hear that they don't need to be that big. What does zfs do with the zil if there is no dedicated zil device? Our servers consist of a small sata drive that holds the OS and a boatload of larger drives on a sas bus. What I'm wondering is if I simply replace the OS disk with an ssd will I get the same performance boost as if I added a dedicated ssd zil? Thanks! Andy On Aug 28, 2012, at 7:07 PM, Josh Paetzel j...@tcbug.org wrote: Original Message Subject: Support for Fusion IO drives? Date: Tue, 28 Aug 2012 16:46:00 -0400 From: Andy Young ayo...@mosaicarchive.com To: freebsd-hardware@freebsd.org We are investigating adding SSDs as ZIL devices to boost our ZFS write performance. I read an article a while ago about iX Systems teaming up with Fusion IO to integrate their hardware with FreeBSD. Does anyone know anything about supported drivers for Fusion IO's iodrives? Thanks! Andy ___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to freebsd-hardware-unsubscr...@freebsd.org I'll put on my iXsystems hat here, as well as my fast storage, ZFS and Fusion-I/O hat. The ZFS filesystem supports dedicated ZIL devices, which can accelerate certain types of write requests, notably related to fsync. The VMWare NFS client issues a sync with every write, and most databases do as well. In those types of environments having a fast dedicated ZIL device is almost essential. In other environments the benefits of a dedicated ZIL range from non-existent to substantial. A good dedicated ZIL device is all about latency. It doesn't need to be large, in fact it will only ever handle 10 seconds of writes, so 10x network bandwidth is worst case. (In most environments this means 20GB is larger than needed). Fusion-I/O cards are far too large to be cost effective ZIL devices. Even though they do rock at I/O latency, the really fast ones are also fairly large, so the $/GB on them isn't so attractive. There are better options for ZIL devices. Another consideration is the Fusion-I/O driver is fairly memory hungry, which competes with memory ZFS wants to use for read caching. Now as an L2ARC device, that's a whole different can of worms. Command line used: iozone -r 4k -s 96g -i 0 -i 1 -i 2 -t 8 Parent sees throughput for 8 readers = 1712399.95 KB/sec L2 ARC Breakdown: 197.45m Hit Ratio: 98.61% 194.71m L2 ARC Size: (Adaptive) 771.13 GiB ARC Efficiency: 683.40m Actual Hit Ratio: 71.09% 485.82m ~ 800GB test data, all served from cache. If you are considering Fusion-I/O, the FreeBSD driver is generally not released to the general public by Fusion-I/O, but can be obtained from various partners. (I believe iXsystems is the only FreeBSD friendly fusion-i/o partner but could be wrong about that) -- Thanks, Josh Paetzel ___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to freebsd-hardware-unsubscr...@freebsd.org ___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to freebsd-hardware-unsubscr...@freebsd.org
Re: Areca vs. ZFS performance testing.
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 8 Jan, 2009, at 02:33 , Danny Carroll wrote: I'd like to post some results of what I have found with my tests. I did a few different types of tests. Basically a set of 5-disk tests and a set of 12-disk tests. I did this because I only had 5 ports available on my onboard controller and I wanted to see how the areca compared to that. I also wanted to see comparisons between JBOD, Passthru and hardware raid5. I have not tested raid6 or raidz2. You can see the results here: http://www.dannysplace.net/quickweb/filesystem%20tests.htm An explanation of each of the tests: ICH9_ZFS5 disk zfs raidz test with onboard SATA ports. ARECAJBOD_ZFS 5 disk zfs raidz test with Areca SATA ports configured in JBOD mode. ARECAJBOD_ZFS_NoWriteCache 5 disk zfs raidz test with Areca SATA ports configured in JBOD mode and with disk caches disabled. ARECARAID 5 disk zfs single-disk test with Areca raid5 array. ARECAPASSTHRU 5 disk zfs raidz test with Areca SATA ports configured in Passthru mode. This means that the onboard areca cache is active. ARECARAID-UFS2 5 disk ufs2 single-disk test with Areca raid5 array. ARECARAID-BIG 12 disk zfs single-disk test with Areca raid5 array. ARECAPASSTHRU_1212 disk zfs raidz test with Areca SATA ports configured in Passthru mode. This means that the onboard areca cache is active. I'll probably be opting for the ARECAPASSTHRU_12 configuration. Mainly because I do not need amazing read speeds (network port would be saturated anyway) and I think that the raidz implementation would be more fault tolerant. By that I mean if you have a disk read error during a rebuild then as I understand it, raidz will write off that block (and hopefully tell me about dead files) but continue with the rest of the rebuild. This is something I'd love to test for real, just to see what happens. But I am not sure how I could do that. Perhaps removing one drive, then a few random writes to a remaining disk (or two) and seeing how it goes with a rebuild. Something else worth mentioning. When I converted from JBOD to passthrough, I was able to re-import the disks without any problems. This must mean that the areca passthrough option does not alter the disk much, perhaps not at all. After a 21 hour rebuild I have to say I am not that keen to do more of these tests, but if there is something someone wants to see, then I'll definitely consider it. One thing I am at a loss to understand is why turning off the disk caches when testing the JBOD performance produced almost identical (very slightly better) results. Perhaps it was a case of the ZFS internal cache making the disks cache redundant? Comparing to the ARECA passthrough (where the areca cache is used) shows again, similar results. -D ___ freebsd...@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-fs To unsubscribe, send any mail to freebsd-fs-unsubscr...@freebsd.org There is a big difference betweeen hardware and ZFS raidz with 12 disk on the get_block test, maybe it would be interesting to rerun this test with zfs prefetch disabled? - -- Regards, Nikolay Denev -BEGIN PGP SIGNATURE- Version: GnuPG v2.0.9 (Darwin) iEYEARECAAYFAkllxT8ACgkQHNAJ/fLbfrnHnwCeJ8nSjBY6fc0Lvu2+fSN5E4HI zb0Ani2ZFLdxYCWYBuCnoo+D244O2lg5 =EKgi -END PGP SIGNATURE- ___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to freebsd-hardware-unsubscr...@freebsd.org
Re: Areca vs. ZFS performance testing.
-BEGIN PGP SIGNED MESSAGE- Hash: SHA1 On 13 Nov, 2008, at 15:59 , Danny Carroll wrote: [snip] It is entirely possible. I do not know however if the Areca cache works just for Raid or also in JBOD mode. I think some RAID controllers do not use the cache when you export the disks as pass-thru/jbod, but on some controllers you can workaround this by making every disk a RAID0(stripe) array with only one disk. Dunno if that would work on the areca... [snip] - -- Regards, Nikolay Denev -BEGIN PGP SIGNATURE- Version: GnuPG v2.0.9 (Darwin) iEYEARECAAYFAkkcQlsACgkQHNAJ/fLbfrkTkgCgo2NupY2Qe3TglJpoIIwne4uH VRwAnRl9p44NFxyWf9zhjrZOOImtiBAs =4Djt -END PGP SIGNATURE- ___ freebsd-hardware@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-hardware To unsubscribe, send any mail to [EMAIL PROTECTED]