My experience converting ~ 400TB worth of PoC1 plots to PoC2 using JohnnyFFM's Blago fork and converter

  • Hi guys,

    Thought I'd share my experience converting my 108 optimized plots (all created using xplotter), into PoC2.

    The JohnnyFFM released are available here.

    The first thing I did was to installed the forked Blago miner and make sure it worked with my existing PoC1 plots. It did.

    Next I converted one of my 2TB plot files inline, to PoC2 while leaving the miner running. As soon as the converter started processing the PoC1 plot, the miner was not able to open that particular plot file, but continued mining as usual on the remaining 107 plots. That was exactly as I expected. Once the plot had been converted to PoC2, the miner picked it up automatically, and was mining away with 107 PoC1 plots and 1 PoC2 plot.

    I next proceeded to load 15 additional 2TB plot files into the first instance of the converter. At this point it has completed 6 of those, and here you can see the average conversion time for those:

    alt text

    As you can see from the above, it takes about 11 hours on average to convert a 2TB plot file inline to PoC2.

    Here's a 8TB plot being converted for comparison:

    alt text

    So it will take about 38 hours to complete, which is about what is expected.

    You can set the amount of RAM each instance of the converter uses. The amount needed, depends on the size of your plot. The following table shows what is needed:

    2TB 1,024 MB
    4TB 2,048 MB
    6TB 3,072 MB
    8TB 4,096 MB

    The amount of RAM you have in your Burst server dictates how many instances of the converter you can run at once. It is all I/O bound and the processor(s) you have doesn't really matter. In my case, I have a server with 64GB of RAM and a pair of E5-2680 V4 14 core CPUs along with 4x LSI HBA's each with 8x6Gbps SAS2 lanes.

    So I'm able to run between 15 and 20 converter instances in parallel, depending on the size of the plots I'm converting at any given time. I have found that the speed of converting each plots does not care about how many instances I'm running at once. It is all about available disk I/O and not processor speed. In my case I have plenty of disk I/O capacity, so my limit is only the amount of RAM. If you run USB connected drives, you probably don't want to do more than a couple of conversions per USB hub.

    Here's a view from the forked Blago miner showing plots that are in the process of being converted (in red), and the ones that have already being converted (in blue):

    alt text

    The plots that are still PoC1 represent 79% in the above pic (as seen in the bottom row).

    Once thing to be aware of is that until PoC2 goes live on main-net at block 500,200, the plot scan speed is reduced by 50%. On the other hand, once we get past block 500,200, scan speed will double.

    On my system, I can scan my 420TB worth of plots in about 22 seconds, to doubling that to 44 seconds until for fork is acceptable to me. Depending on your plot size and the time it takes to scan it, you might not want to convert until we get closer to the hard fork.

    I'm about 70% done converting my plots to PoC2. I anticipate being done by mid week.

    Hope this is useful info for you guys.

  • Conversion completed late last week. Here's a view from the plotter with debug on showing the time it takes to scan each plot/drive. As expected, scan times doubled from what they were before.

    alt text

    Still, 47 seconds ain't bad for 417 TB. Once we hit block 500,200, it will be interesting to see if scan speed improves 4 fold.

  • I converted 1 POC1 plot to POC2 and it actually scans slower than my POC1 plots now. I am only seeing a difference of speed on certain drives for some reason. The plot I used the Conversion tool on was an SAS drive. My POC1 speeds are about the same and now slower on POC2. Guess I don't need to convert! If my speeds go slower on the other drives I will really be in trouble too. I will miss a few quick rounds but will get 50 to 75 percent scanned and still get good submissions. I may rebuild a drive as pure POC2 just to test for you here on SAS.

  • I also just realized I have 2 exact same raid enclosures on USB3 and one scans at 58 seconds for 20TB and the other scans at 102 seconds for all 20TB. One of them was Optimized I think. Can't remember which was which though. They are Both POC1. This might answer why the SAS drives were quicker scans as they would be Optimized. I did these later after Optimization was integrated.

  • admin

    If you're using the modified Blago miner, test out different values for the cache/cache2 values - they can make a huge performance difference.

    I tested with different values from 8K upto 512K. Then tested performance to 50% of plot read and 100% read. This is around 1PB of mixed PoC1 PoC2 plots - around 50/50. Some example readings:

    256K/64K - 50% @ 2,950MB/s, Total: 120 Seconds
    512K/64K - 50% @ 3,120MB/s, Total: 117 Seconds
    512K/8K - 50% @ 3,600MB/s, Total: 189 Seconds (!)
    512K/128K - 50% @ 2,800MB/s, Total: 99 Seconds.

    As the ratio of PoC1 to PoC2 changes, I'd expect the optimal settings to change. As a general rule of thumb it seems to be: Use a larger C1/smaller C2 for a lot of Poc2 files, Large C2/Smaller C1 for a lot of PoC1 files.