How to attain max nonce/min plotting direct to SMR drives



  • So after much consternation about the SMR bottleneck for plotting, I believe I have solved the problem, at least for my configuration.

    As I was watching my plotter, I would see that the Xplotter would run at ~8000 nonce/min for 1/3 of the time then wait 2/3 of the time for the writing to finish. That happened regardless of memory allocation. I tried allocations of 64MB to 8GB always about the same, 1/3 yellow, 2/3 gray. This effectively reduced my plotting speed to 8000/3=2666 n/min.

    I decided to run 3 plots concurrently. The nonces/min will distribute across each of the 3 running programs to add up to ~8000/min.

    If one program goes gray, that computing power is released to the other 2 programs effectively keeping the n/min at max capacity.

    If you notice in the screenshot where there is one gray plotter, the yellow nonces do not add up to ~8000 but I believe this is attributable to the way the n/min is displayed as an average over some time period, when one goes gray the other 2 ramp up to high levels in time. Also there are 2 Jminers running that will slow down the plotter during block scans. This would happen regardless of running 1 plotter or 3.

    Also, I set my Jminers for priority "High" and the 3 plotters to "below normal".

    0_1495206319743_upload-65dd00d1-94b2-4259-b974-36e568ae3570

    0_1495206608693_upload-e4ca316f-d137-4b3f-b0d0-2963b18046c1



  • @rds Interesting notes. For comparison, how many nonces/minute can you get writing to PMR drives? 8,000 nonces/minute might be good for you, but that's too slow for me. It would take roughly 5.5 days to fill up two 8TB drives at that rate. I can easily get 20,000 nonces/minute or better writing to two 8TB PMR drives with a high-end GPU (2.5 days max). An extra day to copy those plots over to SMR drives is still less time than it would take at 8,000 nonces/minute.



  • @sevencardz I recently plotted 8x 8TB Seagate Archive Drives.
    Plot target was a a bunch of fast PMR drives, then merge onto SMR.

    The SMRs were a stripe of 4 drives each. SMR-write-bandwith was >600 MB/s at the beginning, and never slower than 400 MB/s. This was a 32 GiB merge (read 32 GiB, reorder+write 32 GiB, until done).

    But I have to say that this is zfs, a copy-on-write filesystem.
    I have no idea how they perform with filesystems that make heavy use of frequently updated-in-place bitmaps (NTFS, ext[n]).

    Oh, and to give an idea of plot speed: my ancient R9-280x produce 150 MB/s (some 35000 nonces/min), and I run three of those. First stage was ~ 1,2 GByte/s write peak, second stage ~800MB/s read (PMR) and 400-600 MB/s write (SMR).
    So, if you have enough drives 100k nonces/min is no problem.
    <evil grin>



  • @sevencardz said in How to attain max nonce/min plotting direct to SMR drives:

    @rds Interesting notes. For comparison, how many nonces/minute can you get writing to PMR drives? 8,000 nonces/minute might be good for you, but that's too slow for me. It would take roughly 5.5 days to fill up two 8TB drives at that rate. I can easily get 20,000 nonces/minute or better writing to two 8TB PMR drives with a high-end GPU (2.5 days max). An extra day to copy those plots over to SMR drives is still less time than it would take at 8,000 nonces/minute.

    Granted, 8000 may be slow for some but it is the norm for me. Plotting to a PMR alone would still be 8000 but then I have to do the xfer, which take up time, steals resources from other programs and introduces the chance for errors in the file transfer. If a file transfer prematurely aborts, the target file still looks like a full file so you are mining a file that may only have 2% of the nonces in it.

    Try running 2 or 3 gpu plotters, skip the transfer, plot direct and see if you get improved speed.



  • @rds If you have other hardware limitations, then it sounds like you have the best solution for your setup. I've tried writing optimized plots to SMR drives before and the write speed was well below 20 MB/s, which is about what you're getting on each of your drives. That's about the maximum they can handle when doing random writes.



  • @sevencardz , yes but three instances of the plotter is an effective 20x3 MB/s.


Log in to reply
 

Looks like your connection to Burst - Efficient HDD Mining was lost, please wait while we try to reconnect.