Slow snapshots to S3

[jjfalling]

Elasticsearch version: 2.4.2

Plugins installed: [] Cloud AWS / Repository S3

JVM version: OpenJDK 1.8

OS version: Debian 8

Description of the problem including expected versus actual behavior:

Snapshots to S3 are significantly slower then transferring files to S3 using the AWS cli. Uploading a test file using the AWS cli gets ~230mb/s but the snapshots get only around an average of 25mb/s. Both of these tests were preformed on the same EC2 instance. The instance was also in the same region as the S3 bucket. The test index was 10gb with 5 shards and force merged to one segment. The snapshot throttle settings were also set to 10GB.

I also tested ES 5.2.2 and found the snapshot was significantly faster, however still much slower then using the AWS cli.

I've already talked to @tlrx about this issue and we agreed I should open an issue.

[jasontedor]

@tlrx Can you take a look?

[abeyad]

I believe this will be fixed by #24403

[tlrx]

@jjfalling Do you think you could run your tests again but on a version that includes @abeyad 's patch?

[jjfalling]

Sure, I'll try to test it this week.

[jjfalling]

@tlrx I've done some light testing and the snapshots are faster, but still much slower then the AWS cli. I need to do more testing but I'm not sure when I will have time.

[vrizopoulos]

@jjfalling , any progress with this? s3 snapshot speed is still slower in comparison for both backup and restore on 5.6

[tlrx]

@vrizopoulos Yes, we identified a slow down because of a inefficient buffer initialization in the S3 repository and I'll create a fix soon. Also, I'm investigating using a more recent version of the AWS SDK that should speed up uploads (#26993). I'll update this issue once I have more.

[tlrx]

I created #27280 to reduce memory allocation when snapshoting to S3 using Elasticsearch 6.x and I saw a performance improvement. I plan to backport the change down to 6.0 and 5.6.5.

Note that AWS enabled throttling by default in March 2016 so one has to be careful when comparing aws cli and repository-s3 performance because depending of the versions one is throttled and not the other one (also as you noted snapshots are rate limited to 40Mb /sec by default in Elasticsearch).

Also please note that AWS cli and the S3 repository plugin are two different beasts, the former can push files using zero copy send files system calls while the latter has a rate limit by default and does much more than just uploading files.

[vrizopoulos]

@tlrx, Thanks for the updates and your work on this. Based on the AWS article you've linked, it's the number of retries that gets throttled, not necessarily the transfer rate itself?

Regarding the 40Mb/s default limit, is that something that be overwritten when creating the snapshot repo using the "max_snapshot_bytes_per_sec" "max_restore_bytes_per_sec" configuration keys?

Finally is there perhaps a way to allow for multiple restores to execute in parallel that would provide us with a better aggregate performance?

Thanks

[tlrx]

Based on the AWS article you've linked, it's the number of retries that gets throttled, not necessarily the transfer rate itself?

Yes, I linked to this article because @abeyad mentioned the issue #24403 which involves many retries when the incompatible snapshot blob is missing.

Regarding the 40Mb/s default limit, is that something that be overwritten when creating the snapshot repo using the "max_snapshot_bytes_per_sec" "max_restore_bytes_per_sec" configuration keys?

Yes, you can set both of them to 0 or -1 to disable the rate limiter. But test this before changing your production.

Finally is there perhaps a way to allow for multiple restores to execute in parallel that would provide us with a better aggregate performance?

The current snapshot/restore feature does not allow multiple snapshots or restores to be executed at the same time. But for a given snapshot or restore, multiple shards are snapshotted/restored in parallel. If both index-a and index-b were snapshotted at the same time, when restoring the snapshot both indices will be restored in parallel. Does that help?

Finally, I'd love to have more feedback from users about repository-s3 performance on versions 5.6.5/6.0.1/6.1.0 to see the impact of #27280.

[JD557]

Finally, I'd love to have more feedback from users about repository-s3 performance on versions 5.6.5/6.0.1/6.1.0 to see the impact of #27280.

This was an issue that was affecting us on 5.6.4 and the upgrade to 6.1.0 seems to have helped a ton (snapshots are about 30% faster).

Here's a redacted output from _cat/snapshots:

id                   duration indices total_shards
hourly-2018-01-24-11    14.9m     715         3567
hourly-2018-01-24-12    16.1m     717         3577
hourly-2018-01-24-13    15.4m     717         3577
hourly-2018-01-24-14    16.8m     717         3577
------------------------ Update ------------------
hourly-2018-01-24-16    17.4m     717         3577
hourly-2018-01-24-17      11m     717         3577
hourly-2018-01-24-18    11.6m     717         3577
hourly-2018-01-24-19      11m     717         3577
hourly-2018-01-24-20    11.4m     717         3577
hourly-2018-01-24-21    11.6m     717         3577
hourly-2018-01-24-22    10.7m     717         3577
hourly-2018-01-24-23    10.6m     717         3577

[tlrx]

@JD557 Thanks for the feedback!

[eherot]

Is it possible that this might affect deletes as well? I'm running 5.6.3 and snapshot deletes are quite slow:

2018-02-05 06:36:12,600 INFO      Deleting snapshot applogs-2017.10.27...
2018-02-05 07:02:54,618 INFO      Deleting snapshot applogs-2017.10.28...
2018-02-05 07:33:38,949 INFO      Deleting snapshot applogs-2017.10.29...
2018-02-05 07:59:19,782 INFO      Deleting snapshot applogs-2017.10.30...
2018-02-05 08:27:13,283 INFO      Deleting snapshot applogs-2017.10.31...

I don't know exactly what version of ES was used to make these snapshots but I don't think it was < 5.x.