plotman: a Chia plotting manager

This is a tool for managing Chia plotting operations. The tool runs on the plotting machine and provides the following functionality:

• Automatic spawning of new plotting jobs, possibly overlapping ("staggered") on multiple temp directories, rate-limited globally and by per-temp-dir limits.

• Rsync'ing of newly generated plots to a remote host (a farmer/harvester), called "archiving".

• Monitoring of ongoing plotting and archiving jobs, progress, resources used, temp files, etc.

• Control of ongoing plotting jobs (suspend, resume, plus kill and clean up temp files).

• Both an interactive live dashboard mode as well as command line mode tools.

• (very alpha) Analyzing performance statistics of past jobs, to aggregate on various plotting parameters or temp dir type.

Plotman is designed for the following configuration:

• A plotting machine with an array of tmp dirs, a single tmp2 dir, and an array of dst dirs to which the plot jobs plot. The dst dirs serve as a temporary buffer space for generated plots.

• A farming machine with a large number of drives, made accessible via an rsyncd module, and to be entirely populated with plots. These are known as the archive directories.

• Plot jobs are run with STDOUT/STDERR redirected to a log file in a configured directory. This allows analysis of progress (plot phase) as well as timing (e.g. for analyzing performance).

Functionality

Plotman tools are stateless. Rather than keep an internal record of what jobs have been started, Plotman relies on the process tables, open files, and logfiles of plot jobs to understand "what's going on". This means the tools can be stopped and started, even from a different login session, without loss of information. It also means Plotman can see and manage jobs started manually or by other tools, as long as their STDOUT/STDERR redirected to a file in a known logfile directory. (Note: The tool relies on reading the chia plot command line arguments and the format of the plot tool output. Changes in those may break this tool.)

Plot scheduling is done by waiting for a certain amount of wall time since the last job was started, finding the best (e.g. least recently used) tmp dir for plotting, and ensuring that job has progressed to at least a certain point (e.g., phase 2, subphase 5).

Plots are output to the dst dirs, which serve as a temporary buffer until they are rsync'd ("archived") to the farmer/harvester. The archiver does several things to attempt to avoid concurrent IO. First, it only allows one rsync process at a time (more sophisticated scheduling could remove this restriction, but it's nontrivial). Second, it inspects the pipeline of plot jobs to see which dst dirs are about to have plots written to them. This is balanced against how full the dst drives are in a priority scheme.

It is, obviously, necessary that your rsync bandwidth exceeds your plotting bandwidth. Given this, in normal operation, the dst dirs remain empty until a plot is finished, after which it is shortly thereafter picked up by the archive job. However, the decoupling provided by using dst drives as a buffer means that should the farmer/harvester or the network become unavailable, plotting continues uninterrupted.

Screenshot Overview

Plotman 19:01:06 (refresh 9s/20s)  |  Plotting: stagger (1623s/1800s) Archival: active pid 1599918
Prefixes:  tmp=/mnt/tmp  dst=/home/chia/chia/plots  archive=/plots (remote)

  #       plot id    k   tmp   dst    wall   phase    tmp       pid   stat      mem    user    sys     io               
  0   6b4e7375...   32    03   001    0:27     1:2    71G   1590196    SLP     5.5G    0:52   0:02     0s
  1   9ab50d0e...   32    02   005    1:00     1:4   199G   1539209    SLP     5.5G    3:50   0:09     0s
  2   018cf561...   32    01   000    1:32     1:5   224G   1530045    SLP     5.5G    4:46   0:11     2s
  3   f771de9c...   32    00   004    2:03     1:5   241G   1524772    SLP     5.5G    5:43   0:14     2s
...
 16   58045bef...   32    10   002   11:23     3:5   193G   1381622    RUN     5.4G   15:02   0:53   0:02
 17   8134a2dd...   32    11   003   11:55     3:6   148G   1372206    RUN     5.4G   15:27   0:57   0:03
 18   50165422...   32    08   001   12:43     3:6   102G   1357782    RUN     5.4G   16:14   1:00   0:03
 19   100df84f...   32    09   005   13:19     4:0      0   1347430    DSK   705.9M   16:44   1:04   0:06

tmp   ready    phases     tmp   ready    phases        dst   plots   GB free         phases         priority 
 00      --   1:5, 3:4     06      --   2:4            000   1       1890      1:5, 2:2, 3:4        47
 01      --   1:5, 3:4     07      --   2:2            001   0       1998      1:2, 1:7, 3:2, 3:6   34
 02      --   1:4, 3:3     08      --   1:7, 3:6       002   0       1967      1:6, 2:5, 3:5        42
 03      --   1:2, 3:2     09      --   2:1, 4:0       003   0       1998      1:6, 3:1, 3:6        34
 04      OK   3:1          10      --   1:6, 3:5       004   0       1998      1:5, 2:4, 3:4        46
 05      OK   2:5          11      --   1:6, 3:6       005   0       1955      1:4, 2:1, 3:3, 4:0   18

Archive dirs free space
000:   94GB | 005:   94GB | 012:   24GB | 017:   99GB | 022:   94GB | 027:   94GB | 032: 9998GB | 037: 9998GB
001:   94GB | 006:   93GB | 013:   25GB | 018:   94GB | 023:   94GB | 028:   94GB | 033: 9998GB |
002:   93GB | 009:   25GB | 014:   93GB | 019:   31GB | 024:   94GB | 029: 7777GB | 034: 9998GB |
003:   94GB | 010:   25GB | 015:   94GB | 020:   47GB | 025:   94GB | 030: 9998GB | 035: 9998GB |
004:   94GB | 011:   25GB | 016:   99GB | 021:   93GB | 026:   94GB | 031: 9998GB | 036: 9998GB |

Log:
01-02 18:33:53 Starting plot job: chia plots create -k 32 -r 8 -u 128 -b 4580 -t /mnt/tmp/03 -2 /mnt/tmp/a -d /home/chi
01-02 18:33:53 Starting archive: rsync --bwlimit=100000 --remove-source-files -P /home/chia/chia/plots/004/plot-k32-202
01-02 18:52:40 Starting archive: rsync --bwlimit=100000 --remove-source-files -P /home/chia/chia/plots/000/plot-k32-202

The screenshot shows some of the main features of Plotman.

The first line shows the status. The plotting status shows whether we just started a plot, or, if not, why not (e.g., stagger time, tmp directories being ready, etc.; in this case, the 1800s stagger between plots has not been reached yet). Archival status says whether we are currently archiving (and provides the rsync pid) or whether there are no plots available in the dst drives to archive.

The second line provides a key to some directory abbrevations used throughout. For tmp and dst directories, we assume they have a common prefix, which is computed and indicated here, after which they can be referred to (in context) by their unique suffix. For example, if we have tmp dirs /mnt/tmp/00, /mnt/tmp/01, /mnt/tmp/02, etc., we show /mnt/tmp as the prefix here and can then talk about tmp dirs 00 or 01 etc. The archive directories are the same except that these are paths on a remote host and accessed via an rsyncd module (see config.yaml for details).

The next table shows information about the active plotting jobs. It is abbreviated to show the most and least recently started jobs (the full list is available via the command line mode). It shows various information about the plot jobs, including the plot ID (first 8 chars), the directories used, walltime, the current plot phase and subphase, space used on the tmp drive, pid, etc.

The next tables are a bit hard to read; there is actually a tmp table on the left which is split into two tables for rendering purposes, and a dst table on the right. The tmp tables show the phases of the plotting jobs using them, and whether or not they're ready to take a new plot job. The dst table shows how many plots have accumulated, how much free space is left, and the phases of jobs that are destined to write to them, and finally, the priority computed for the archive job to move the plots away.

The last table simply shows free space of drives on the remote harverster/farmer.

Finally, the last section shows a log of actions performed -- namely, plot and archive jobs initiated. This is the one part of the interactive tool which is stateful. There is no permanent record of these executed command lines, so if you start a new interactive plotman session, this log is empty.

Limitations and Issues

The system is tested on Linux only. Plotman should be generalizable to other platforms, but this is not done yet. Some of the issues around making calls out to command line programs (e.g., running df over ssh to obtain the free space on the remote archive directories) are very linux-y.

The interactive mode uses the curses library ... poorly. Keypresses are not received, screen resizing does not work, and the minimum terminal size is pretty big.

Plotman assumes all plots are k32s. Again, this is just an unimplemented generalization.

Many features are inconsistently supported between either the "interactive" mode or the command line mode.

There are many bugs and TODOs.

Installation

This program requires psutil, pyfakefs, and texttable.

Installation for Ubuntu 20.04:

# Install Python package system
sudo apt-get install python3-pip

python3 -m pip install psutil
python3 -m pip install pyfakefs
python3 -m pip install texttable