How to measure Bus load



  • @IncludeBeer well, I have no idea.



  • @vaxman said in How to measure Bus load:

    @IncludeBeer said in How to measure Bus load:

    I'm expanding my farm, and now have 3 drives plotting on the same CPU using XPlotter. I also mine on this same computer (12 drives). I noticed in the last few drives I added, the write speed on my drives will plummet after just a couple iterations from xplotter.

    what does plummet mean ? They go at 150 MB/s first and then drop to 10 ?
    PMR or SMR drives ?

    The 150 would indicate cache writing then the real world speed kicks in.



  • @rds ..and the real world speed is 10 MB/s on a PMR drive ?



  • @vaxman said in How to measure Bus load:

    @rds ..and the real world speed is 10 MB/s on a PMR drive ?

    Not saying it's right but for now real :). I only use 1-2 GB for plotting so the chunks sent to the disk are small, my write speed runs up and runs right back down, as the chunk transfer is completed. For plotting the goal is to keep the counter line yellow, then the displayed Nonces/min is as fast as you can go.



  • Well, these particular drives are rated at 120mb/s "max xfer speed". Suffice to say, 2-20 is certainly low.

    I'll run some experiments with crystaldisk, both while plotting & mining (bus is in constant use), and with a completely idle system. Will post back the results...



  • Ok, here are my results. I ran all CrystalDisk tests, with 5 runs on each test, on the same 4tb hard drive, connected to the same USB 3.0 port with same USB 3.0 cable.

    While mining Burst (12 drives) AND plotting 2 external drives:
    alt text

    Completely idle system:
    alt text

    According to these results, write speed is not affected at all by mining/plotting (meaning it's safe to assume my BUS isn't the bottleneck).


  • Mod

    XPlotter uses random write to plot - threads*64Kb per once, see "4K" in your tests. (if XPlotter used 8 threads - see "512K" in tests)
    Miner uses sequential reads from plot - "Seq" in tests



  • @Blago Thanks Blago, that explains it!



  • @Blago said in How to measure Bus load:

    XPlotter uses random write to plot - threads*64Kb per once, see "4K" in your tests. (if XPlotter used 8 threads - see "512K" in tests)
    Miner uses sequential reads from plot - "Seq" in tests

    I don't get what you're saying? IIRC a nonce is 262144 bytes. So if 4 threads then it would be 65536, if 8 threads it would be 32768, why are you saying the 4k test is representative of the 4 threads and the 512k?? test is representative of the 8 thread condition?


  • Mod

    @rds yes, but plotter writes 64 bytes (scoop) of one nonce, then seek to another place of file, writes another 64 bytes....
    In brief, each thread calculates matrix [64*4096][1024] and then write to file [64][1024] * 4096 times
    https://github.com/Blagodarenko/XPlotter/blob/master/XPlotter.cpp#L107
    https://github.com/Blagodarenko/XPlotter/blob/master/Nonce.cpp#L130

    why 1024 nonces (64kb for write per once of each threads) by default?
    writing time of 6 * 800 nonces ~= 45 sec
    writing time of 6 * 1024 nonces ~= 45 sec
    writing time of 6 * 1030 nonces ~= 90 sec
    writing time of 6 * 2048 nonces ~= 90 sec
    64Kb it's optimal size
    http://zabkat.com/blog/buffered-disk-access.htm (last column)


Log in to reply
 

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