[LU-16613] Future: folio_batch BIO write path

• Status: Open
• Project: Lustre
• Type: Improvement
• Priority: Minor
• Reporter: Shaun Tancheff
• Assignee: Shaun Tancheff
• Resolution: Unresolved

Description

Prototype support for folio_batch support on page cache backed buffer I/O path.

Relies on a patched kernel to provide:

• generic_perform_batch_write
• aops->write_batch_begin and aops->write_patch_end
• filemap_dirty_folio_batched
• grab_cache_folios_fast

Comments

[Gerrit Updater]

"Shaun Tancheff <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/50194
Subject: LU-16613 clio: Use folio_batch write path
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 6dc4f8a03385bd507f5ebb5d67503c10e05053c7

[Patrick Farrell]

Shaun,

This is very exciting, I didn’t know the requisite folio stuff existed yet - have you been able to benchmark this at all?  And is there some form of read batch available as well?  The read path is also stuck in page cache work as its main performance limitation.

[Patrick Farrell]

Ah, I see it requires a patched kernel.  Can you say more about this in general? Where are these patches from, are they expected upstream, etc, etc

[Shaun Tancheff]

I have patches here:
https://github.com/stancheff/linux/
branch: dedupe-v1

The basic idea is to enable 'batch' adding to the page cache (and batch dirty pages). I did the first pass on the write path.

There is a device mapper that can be used to do local perf testing, ex:

sudo mkdir -p /vol/fcp 
echo "0 `sudo blockdev --getsz /dev/vda` dedupe /dev/vda 0" | sudo dmsetup create dedupe
sudo mkfs.ext4 -F -E 'lazy_itable_init=0 packed_meta_blocks=1' /dev/mapper/dedupe
sudo mount /dev/mapper/dedupe /vol/fcp
sudo fio --filename=/vol/fcp/fio.bin \
  --size=100GB \
  --buffered=1 \
  --rw=write \
  --bs=64k \
  --buffer_pattern=0x00000000 \
  --ioengine=sync \
  --iodepth=64 \
  --runtime=120 \
  --numjobs=16 \
  --time_based \
  --group_reporting \
  --name=throughput-test-job \
  --eta-newline=1

Locally I see 2 G/s buffered. Still a long ways from direct I/O which is closer to 12G/s

[Shaun Tancheff]

The read path is batch'd:

 generic_file_read_iter() 
    filemap_read()
      filemap_get_pages()