JorgeB Posted March 4, 2016 Share Posted March 4, 2016 Diskspeed graph below, 75MB/s at 70% makes sense, half its speed, but the same speed or a little more in the beginning makes me believe I’m hitting some limitation, don’t have a clue what it is. Parity check with these disks closely follows the graph. Quote Link to comment
bobkart Posted March 5, 2016 Author Share Posted March 5, 2016 Yeah that graph definitely rules out my potential explanation for what you're seeing. Interesting. Your numbers using SSDs would seem to disprove any theory that it's the operating system, or motherboard, or disk controllers. That leads me to start thinking along the lines of it having something to do with rotational latency. But I'm way out of my areas of expertise on this, it's just the only thing I can think of so far that might prevent effective write speeds from scaling with increased drive *sequential* throughput. With SSDs not suffering from rotational latency, they would exhibit effective write speeds much more reflective of their sequential throughput. Again just reaching for *something* in the way of an explanation for what you're seeing, I could easily be way off. In addition to trying various md_write_limit values, did you also try the so-called Turbo Write Mode (md_write_method = 1)? Quote Link to comment
JorgeB Posted March 5, 2016 Share Posted March 5, 2016 With turbo write on I max out gigabit. I also think it may have something to do with rotational latency, the 8TB Seagate has a very similar performance curve, so in theory it should perform about the same, but if rotational latency is a factor it should be noticeable slower, if I have time I’ll do some tests this weekend. Quote Link to comment
opentoe Posted March 5, 2016 Share Posted March 5, 2016 With turbo write on I max out gigabit. I also think it may have something to do with rotational latency, the 8TB Seagate has a very similar performance curve, so in theory it should perform about the same, but if rotational latency is a factor it should be noticeable slower, if I have time I’ll do some tests this weekend. I shouldn't complain anymore. I get 140MB/sec average on my keeper drives and they are all %75 or more full. Once we get into the 8-12TB land it will just take too much to move that amount of data with the controllers we have today. I guess once that data is on that 8TB drive it is staying there. Period. Too bad we all can't afford some nice enterprise grade stuff. Quote Link to comment
JorgeB Posted March 5, 2016 Share Posted March 5, 2016 I get 140MB/sec average on my keeper drives and they are all %75 or more full. 140MB/s write speed?? Quote Link to comment
opentoe Posted March 6, 2016 Share Posted March 6, 2016 I get 140MB/sec average on my keeper drives and they are all %75 or more full. 140MB/s write speed?? Not on the array, just with writing to a single drive. Quote Link to comment
JorgeB Posted March 6, 2016 Share Posted March 6, 2016 I did some tests on my server with 8TB Seagates, did two copys: to disk3 (38% full): ~60MB/s (disk speed at this position ~175MB/s) to disk1 (97% full): ~50MB/s (disk speed at this position ~100MB/s) Based on these and the Toshiba results I’m thinking that rotational latency does limit max write speed in Unraid and 5400/5900rpm disks max out at ~60MB/s and 7200rpm disks at ~80MB/s. Quote Link to comment
bobkart Posted March 6, 2016 Author Share Posted March 6, 2016 You're definitely on to something JB. Once I hit a week straight of stable uptime on my new server (another day and a half to go), I'll be adding another 8TB drive, then I'll be in a position to test write speeds to an empty one of these drives. I still wonder if motherboard/disk controllers might have something to do with it, not to the point of even slowing down SSDs, but enough to cause the rotational latency to be a problem, versus some other configuration that lets them be written to faster. I guess we'll only know if that's the case if/when a counterexample is produced, where someone is able to write to these drives faster than ~60MB/s. I'm hoping to be that counterexample, but am not really expecting to be. Can you share details of the motherboard/disk controllers used in your test? Quote Link to comment
JorgeB Posted March 8, 2016 Share Posted March 8, 2016 Toshiba tests were done on a Supermicro X9SCM-F using the onboard controller, the Seagate were on a HP N54L microserver also using the onboard controller, both in AHCI mode. Note that both servers can write at 100MB/s+ using the same controller with turbo writes on, this mode need considerably more bandwidth than the normal writing mode, so I believe the controllers are not the reason for the limits. Quote Link to comment
bobkart Posted March 8, 2016 Author Share Posted March 8, 2016 Yeah it wouldn't be so much the raw throughput of the controllers (your SSD test confirmed that's not a problem) as some timing issues that, combined with the drives' rotational latency and unRAID's read-modify-write technique, result in the behavior you're seeing. Back to the build: I hit a solid week of uptime with zero issues. I'll be adding a 3-bay hot swap drive cage next; I anticipate about a 3-watt hit for that. Just started a Parity Check, after first capturing the tunables I used last time in a script, so I can easily reapply them. This is because I see md_write_limit being set to a different value in disk.cfg, and md_sync_thresh isn't even listed. Five minutes in and I'm showing over 200MB/s, so far so good . . . Quote Link to comment
bobkart Posted March 11, 2016 Author Share Posted March 11, 2016 UPDATE Parity Check finished in 16 hours 10 minutes, within 3.5 minutes of the previous run. I had increased the poll_spindown frequency from 1/1800 to 1/60, so that might be related. On the power consumption front: starting with 29.0 watts idle, I replaced the 2x 16GB ECC UDIMMs with 4x 8GB ECC UDIMMs (I needed the 16GB ECC UDIMMs on another Avoton board). Idle power consumption went up to 29.6 watts. Then I added the three-bay hot-swap drive cage, idle power consumption went up to 30.8 watts. So now I have nine-drive capacity, although I've yet to enable the other six SATA ports. Here is the drive cage: iStarUSA BPN-DE230SS-BLACK 2 x 5.25" to 3 x 3.5" SAS/SATA Trayless Hot-Swap Cage - OEM http://www.newegg.com/Product/Product.aspx?Item=N82E16816215240 I'll be adding another data drive at some point. The plan is that the parity drive and a warm spare drive will be in the hot-swap cage, plus one extra bay for a guest drive. If/when dual parity becomes available, that warm spare drive will turn into the second parity drive. Quote Link to comment
JorgeB Posted March 11, 2016 Share Posted March 11, 2016 That’s a good average speed but with some tuning you could probably shave an hour, these are from an HP N54L with 4 x 8TBs, before and after tuning: 16h46m13s – 132.5MB/s 15h01m14s – 148MB/s Quote Link to comment
bobkart Posted March 11, 2016 Author Share Posted March 11, 2016 That’s a good average speed but with some tuning you could probably shave an hour, these are from an HP N54L with 4 x 8TBs, before and after tuning: 16h46m13s – 132.5MB/s 15h01m14s – 148MB/s Nice improvement for sure. I actually typo'd that last post, my times are all in the 15-hour range . . . you can see my before-and-after-tuning improvement, where I pick up around half an hour (~142MB/s -> ~147MB/s) Quote Link to comment
JorgeB Posted March 11, 2016 Share Posted March 11, 2016 I actually typo'd that last post, my times are all in the 15-hour range . . . you can see my before-and-after-tuning improvement, where I pick up around half an hour (~142MB/s -> ~147MB/s) OK, that’s very close to optimal for these disks, in that case you’re good. Quote Link to comment
bobkart Posted March 19, 2016 Author Share Posted March 19, 2016 UPDATE Found that the Intel SATA ports support hotswap whereas the Marvell SATA ports don't appear to. So enabled all the Marvell SATA ports, switched the six 8TB drives to all connect using those ports, and ran a Parity Check. The Parity Check completed in one second less time that with the six 8TB drives connected via the Intel SATA ports. Power consumption is at 32.1W idle now, ready to connect more drives via the three-bay hotswap cage. Quote Link to comment
bobkart Posted March 30, 2016 Author Share Posted March 30, 2016 I swapped out the 4x 8GB ECC UDIMMs for 2x 16GB ECC UDIMMs . . . power consumption dropped back down under 32 watts. Also upgraded to unRAID 6.1.9. Connected the three-bay hotswap cage to the motherboard (had to wait for 18" versions of those StarTech SATA cables to get here, 12" was not enough). Just plugged in a small drive, with the system powered up and the array started, to test the hotswap capability of the SATA ports. No problem, drive recognized, doing a preclear on it now. I'm at 96% capacity now so I'll be adding another data drive soon. Once the dust settles on that, I'll be thinking about upgrading to 6.2.0, I see there's a beta20 release now; I have beta19 on another server and it's working well so far. Quote Link to comment
garycase Posted March 30, 2016 Share Posted March 30, 2016 Not only did that change help your power consumption, but using 2 modules in an unbuffered system will result in a MUCH "cleaner" and more reliable waveform for the signaling on the memory bus than 4 modules did ... so you've also improved the reliability of your system Quote Link to comment
bobkart Posted April 19, 2016 Author Share Posted April 19, 2016 Started seeing "kernel: BUG: unable to handle kernel paging request at ..." today in the log. Every 17 or so minutes for several hours. Rebooted, running Parity Check, log is clean so far, over an hour in. Searches of the forums seem to indicate that this is a potential memory corruption problem. And difficult to track down (that part I know, I develop software myself). If anyone else has seen this or know what might be causing it, I'd be interested to hear about it. In the meantime I'll be keeping an eye out for this showing up again. Quote Link to comment
bobkart Posted May 19, 2016 Author Share Posted May 19, 2016 30+ days later and no sign of the apparent kernel bug. Still holding steady at 32 watts (idle) for 40TB net capacity; the three hotswap bays remain empty. I have 2TB unused capacity, so it will be a little while before I add another data drive . . . most likely I'll move to Dual Parity at that time. Then adding a warm spare (as unused Cache drive) will round out the build: six data drives, two parity drives, one warm spare, with the data drives inside the chassis and the spare/parity drives in the hotswap bays. I believe I'll be under 36 watts idle in that configuration, so I should be back under 3/4 watt per protected TB. Quote Link to comment
bobkart Posted September 1, 2016 Author Share Posted September 1, 2016 UPDATE - Lots of changes: increased RAM from 32GB to 64GB increased protected capacity from 40TB to 48TB by adding 6th data drive upgraded to unRAID 6.2.0-beta21 (will go to rc4 or later at my next power cycle) added second Parity drive added warm spare (as unused Cache drive) Power consumption at idle hovers in the mid-34-watt range. So easily under 35 watts. Result is about 0.72 watts idle per protected terabyte. I welcome people posting their capacity/power numbers here. I understand that energy efficiency is not everyone's priority; still, if anyone has impressive numbers, I'd be interested to see them. Quote Link to comment
bobkart Posted August 15, 2017 Author Share Posted August 15, 2017 Went to write some files to this server earlier today and it was unresponsive. No video, wouldn't boot. Motherboard looks to have failed. I just finished moving the drives to a hastily-assembled system, using parts I had laying around. No data loss as far as I can tell. Got everything backed up and am running a parity check now. By coincidence I had just ordered another motherboard, to start a build of an online backup (mirror) system for this one. Looks like that new board (ASRock Rack C236 WSI) will instead replace this failed Avoton board. Quote Link to comment
garycase Posted August 16, 2017 Share Posted August 16, 2017 I presume you tried another power supply ... "just in case" "No video" makes this unlikely -- but are you sure it's not just hanging due to a switch in the boot device (if it's attempting to boot from a hard drive instead of the USB flash drive it will generally hang with a blank screen. This can happen after a power failure ... especially if the motherboard battery has failed. But you should still be able to boot to the BIOS, so assuming you have a keyboard and display connected, the "no video" comment likely excludes this. Just wanted to be sure you'd considered it. Quote Link to comment
bobkart Posted August 16, 2017 Author Share Posted August 16, 2017 Thanks for the advice Garycase. I tested the power supply with a power-supply tester, so it's not 'obviously' failed, but could still be off just enough to prevent that board from booting. Will circle back for that. No video comes out at any time over the course of what would be the boot sequence. Everything powers up (drives, fans) and LEDs on the motherboard come on, but no video. That might be acceptable for now it if would finish the boot sequence, but no joy. I also cycled through the DIMMs, one at a time in the primary DIMM slot, to rule out a DIMM failure. I also reset the CMOS (by removing the battery for several minutes). I'm using a UPS so should not have experienced any power loss. There was a series of serious-looking errors in the log (I was 'tail -f'-ing the log in Emacs/Cygwin). Something about I/O and locking I think. Was hoping a reboot cleared it up, so I didn't save it. Quote Link to comment
HellDiverUK Posted August 23, 2017 Share Posted August 23, 2017 Ah, good old AsRock. Clearly they don't even bother checking their boards before stuffing it in the box. I got an expensive AsRock Workstation board not too long ago, factory sealed, and the socket was totally mangled. Quote Link to comment
bobkart Posted August 25, 2017 Author Share Posted August 25, 2017 (edited) New motherboard and processor is in place, working well so far. WIll run a Parity Check soon. Only have 4GB of non-ECC memory in there for now, but the change gave a substantial improvement in power consumption: 23.0 watts at idle, compared to mid-33's before. *And* more processing power (i3-6100T). That's over 2TB of single-parity-protected storage per idle watt. It will creep up more towards 24 watts when I put more memory in. Edited August 25, 2017 by bobkart Quote Link to comment
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.