s3 performance is slow

[tyommik]

Bug Report

s3 performance is slow

Description

Ticket is based on topic in discord (“need-help” channel).
The problem is that I tried to use DVC with miniO (s3 compatible storage) and noticed that its performance is very slow.

My env:

• miniO storage location is on SSD (Intel Optane, high performance)
• MiniO server <------ 1000Mbit/sec --------> DVC (s3 client)
• Bucket 40Gb - 410k files (each file <= 400KB)

DVC version: 2.0.17 (deb)
---------------------------------
Platform: Python 3.8.8 on Linux-5.4.0-70-generic-x86_64-with-glibc2.4
Supports: All remotes
Cache types: hardlink, symlink
Cache directory: ext4 on /dev/sda
Caches: local
Remotes: s3
Workspace directory: ext4 on /dev/sda
Repo: dvc, git

When I did dvc pull -j 20 maximum speed was 80 mbit/sec but average was about 40. I download the bucket for 220 minutes.

What I tried else:
dvc pull -j 80 - no improvements.
awscli - aws tool can maximum 160 Mbit/sec downloading speed. I tried different settings, but I couldn't exceed the limit.
s4cmd - maximum I got is about 130 Mbit/sec, 64minutes to get the bucket.
s5cmd - maximum 960 Mbit/sec and less than 10 minutes to download the whole bucket (GoLang)

So you can see that storage performance is okay but the download speed of tools written on python can not reach maximum.

Reproduce

Profiler stat: https://disk.yandex.ru/d/XNajwHgWYlPSHA

[pared]

Related iterative/dvc#5683

[diehl]

FYI I'm seeing very slow transfer speeds to S3 when doing a dvc push. In comparison, when doing a aws s3 cp, I'm seeing throughput that is about 10-100x faster from the same machine.

Here's the output from dvc doctor on my machine.

DVC version: 2.45.0 (brew)
--------------------------
Platform: Python 3.11.2 on macOS-12.6.3-x86_64-i386-64bit
Subprojects:
	dvc_data = 0.40.1
	dvc_objects = 0.19.3
	dvc_render = 0.1.2
	dvc_task = 0.1.11
	dvclive = 2.0.2
	scmrepo = 0.1.9
Supports:
	azure (adlfs = 2023.1.0, knack = 0.10.1, azure-identity = 1.12.0),
	gdrive (pydrive2 = 1.15.0),
	gs (gcsfs = 2023.1.0),
	http (aiohttp = 3.8.4, aiohttp-retry = 2.8.3),
	https (aiohttp = 3.8.4, aiohttp-retry = 2.8.3),
	oss (ossfs = 2021.8.0),
	s3 (s3fs = 2023.1.0, boto3 = 1.24.59),
	ssh (sshfs = 2023.1.0),
	webdav (webdav4 = 0.9.8),
	webdavs (webdav4 = 0.9.8),
	webhdfs (fsspec = 2023.1.0)
Cache types: reflink, hardlink, symlink
Cache directory: apfs on /dev/disk3s1s1
Caches: local
Remotes: s3
Workspace directory: apfs on /dev/disk3s1s1
Repo: dvc, git

[efiop]

@diehl Are you also using minio?

[diehl]

@efiop I'm not. I don't know what minio is.

[efiop]

@diehl So real aws s3 then, right?

[diehl]

@efiop that is correct. using the AWS CLI.

[efiop]

@diehl Btw, what does the data that you have look like? Whole directories tracked by dvc? How many files there typically, what approx size? IIRC you were talking about geojson before, so I suppose thousands of ~100M files?

[efiop]

Chatted with @diehl privately. For the record, there are two directories, one 5G total with 224 files and the other is 14G total with 63 files. Inside there are misc files, with 5G being the biggest one in the second dataset. So likely we need to look at a scenario of transferring individual large files.

Need to reproduce locally and see if there is anything on the surface.

[pmrowla]

This is probably an s3fs issue similar to the adlfs one we had recently.

s3fs _put_file()/_get_file() do support chunked (multipart) upload/downloads, but chunks within a single file are always done in sequence rather than in parallel. s3fs should probably be using asyncio.gather() or fsspecs run_coros_in_chunks() to batch the chunk transfers, and then reassemble them at the end as needed. I'm guessing this hasn't been addressed because doing the chunk transfers in parallel requires balancing those batched tasks with the file-level batching.

[efiop]

@pmrowla Do you think we should prioritize it and solve it? What would be your estimation there?

[pmrowla]

Handling it the same way we did with adlfs (where we only optimize the worst-case single large file upload/download scenario) should be relatively quick to do in s3fs, but I should note that it looks like they have tried to do concurrent multipart uploads before in some of the s3fs calls but ran into problems that made them revert to only sequential operations: https://github.com/fsspec/s3fs/blob/b1d98806952485be86379f0f4574ee4de24568a1/s3fs/core.py#L1768C31-L1769

[dberenbaum]

It looks like it multipart upload requires a minimum size of 5 MiB for all but the last part: https://docs.aws.amazon.com/AmazonS3/latest/userguide/qfacts.html

Also, to set expectations, the azure approach currently will only help when doing dvc push on a single file right @pmrowla?

[pmrowla]

Also, to set expectations, the azure approach currently will only help when doing dvc push on a single file right @pmrowla?

Yes that's correct

[pmrowla]

Should be resolved by fsspec/s3fs#848

There's some raw s3fs numbers in the upstream PR, but for reference with DVC and a single file that takes aws s3 cp ~40 seconds to upload:

with s3fs main:

time dvc push -r s3-us
Collecting                                                                                                                                                                     |1.00 [00:00,  952entry/s]
Pushing
1 file pushed
dvc push -r s3-us  10.49s user 4.12s system 13% cpu 1:51.14 total

with s3fs PR and default --jobs (40 on my machine):

time dvc push -r s3-us
Collecting                                                                                                                                                                     |1.00 [00:00,  930entry/s]
Pushing
1 file pushed
dvc push -r s3-us  7.14s user 3.01s system 27% cpu 36.642 total

[pmrowla]

This is merged upstream and will be available in the next s3fs release

[diehl]

@pmrowla do you have an ETA for the next s3fs release by chance?

[pmrowla]

I don't have an ETA but generally the fsspec/s3fs maintainers are fairly quick about doing releases for work we've contributed upstream

cc @efiop

[diehl]

Roger that - thanks @pmrowla