Q: `pod5 subset <..> --threads` takes a lot of resources

[sklages]

I just started using pod5 subset to regroup my POD5 data per-channel to optimize dorado duplex basecalling speed.

The default setting for --threads is 4, I was brave and set it to 8.. well, ..

System: Linux, 128 cores, 1TB RAM

Load before starting pod5 was around 34 ..

Quickly after starting sth like:

    pod5 subset \
        --threads 8 \
        --force-overwrite \
        --recursive \
        --summary $rc_file \
        --columns channel \
        --output $POD5_TMPDIR \
        $RAW_DATA_DIR

with a small (7Gbyte) P2 dataset, the system load went over 200, making the system a bit sluggish. Heavy I/O I guess .. top showed the 8 python processes, each with 700 to 1200% in different states (R,S,D) ...

top - 21:43:28 up 274 days, 10:38, 29 users,  load average: 214.81, 106.81, 77.11
<...>
   PID USER      PR  NI    VIRT    RES    SHR S  %CPU  %MEM     TIME+ COMMAND
 19971 USER      20   0   22.2g 349288  51752 R  1210   0.0  36:49.68 python3
 19972 USER      20   0   21.6g 288420  51156 S  1041   0.0  37:35.33 python3
 19967 USER      20   0   21.7g 377372  50984 D 989.2   0.0  36:09.18 python3
 19969 USER      20   0   21.7g 277636  48848 R 972.3   0.0  35:51.98 python3
 19974 USER      20   0   21.7g 380280  54068 S 949.0   0.0  38:09.10 python3
 19973 USER      20   0   21.7g 362636  48292 R 943.6   0.0  35:44.70 python3
 19968 USER      20   0   21.6g 286052  48144 S 875.5   0.0  37:15.61 python3
 19970 USER      20   0   21.7g 270932  47140 D 843.0   0.0  36:15.94 python3

This is by far more than I would expect when using --threads 8 ...

Is there a smooth way to better control the CPUs used by pod5 subset and thus I/O throughput (avoiding sth like taskset)? This small dataset took ~7 minutes to finish, the larger datasets are 100x to 200x in size. So I wonder what is considered "best practice"? On shared servers this is quite a big issue.

Instead of writing a few thousand per-channel POD5 files, wouldn't it be convenient to write one or some more large(r) per-channel-sorted POD5 files?

Any ideas/comments/remarks are welcome :-)

[sklages]

Now using a large P2 dataset but leaving the --threads on default (4).

This runs out of memory on the same machine, 1TB RAM (~550G free), reformatted for better readability:

### [2023-12-02 07:27:45] START: Merging POD5 files by channel..
Parsed 51501778 targets
memory allocation of 2883584000 bytes failed

(core dumped) pod5 subset 
  --threads 4 
  --force-overwrite 
  --recursive 
  --summary $rc_file 
  --columns channel 
  --output $POD5_TMPDIR 
  $RAW_DATA_DIR

Both RAW_DATA_DIR and POD5_TMPDIR are not on the same storage.

$rc_file is 2GB in size,

$ head pod5_summary_per-channel.tsv
read_id channel
7636a648-a348-4b3a-9925-39587c4dfbbd    2727
0355921b-3da6-45a7-af50-45b9a7e1d93f    1940
bf818b42-7108-4f63-9115-0df5a66c9db6    1572
<...>

$ pod5 -v
Pod5 version: 0.3.2

Is there something very obvious which I am missing here?

[HalfPhoton]

Hi @sklages , we are in the process of updating subset to use significantly fewer resources.

Thanks for raising this issue. We'll let you know when we push these changes up.