Bug #5901

[API] Improve performance of large requests in parallel

Added by Brett Smith over 2 years ago. Updated 8 months ago.

Status:NewStart date:05/04/2015
Priority:NormalDue date:
Assignee:-% Done:

0%

Category:API
Target version:Arvados Future Sprints
Story points-
Velocity based estimate-

Description

Attached are two files. The first is a simple Python script that uses threads to fetch the same collection object from the API server multiple times simultaneously. Currently, the collection's manifest is 75492690 bytes. The collection UUID is su92l-4zz18-wd2va9q9lnfx6ga

The log file was generated by running:

for n in 2 4 6 8; do python multi.py "$n" || break; done | tee multi.log

Simply, it shows that performance takes a noticeable dive as the number of simultaneous requests increase. The eight-thread calls never succeed; instead they raised a timeout exception. This problem just bit a real user: parallelizing over many files in this collection, the first batch of parallel tasks all failed because they all tried to fetch the collection simultaneously, and timed out waiting for an API server response. We have to improve performance here to make sure this use pattern doesn't fail.

multi.log - Log file showing degraded performance (1.41 KB) Brett Smith, 05/04/2015 08:25 pm

multi.py Magnifier - Test script to demonstrate issue (691 Bytes) Brett Smith, 05/04/2015 08:25 pm

multi1.log (1.74 KB) Tom Morris, 02/22/2017 11:40 pm


Related issues

Related to Arvados - Bug #5902: [Workbench] collection#show is unresponsive for a large c... New 05/04/2015

History

#1 Updated by Brett Smith over 2 years ago

#2 Updated by Brett Smith over 2 years ago

  • Target version changed from Bug Triage to Arvados Future Sprints

#3 Updated by Tom Morris 8 months ago

This has improved by over an order of magnitude(!) since 2015 which is great, but 20 seconds to fetch 75 MB still seems like an awful lot of time and a 3-4x stretch factor under an 8x load when the data should already be cached also seems out of line.

Threads Elapsed (2015) Elapsed (2017)
2 275.6 20.8
2 283.0 25.0
4 285.3 28.5
4 287.2 31.7
4 390.5 33.9
4 396.0 53.7
6 654.8 28.6
6 919.5 35.4
6 923.7 37.2
6 931.3 49.5
6 933.5 75.7
6 934.4 86.2
8 - 38.8
8 - 49.3
8 - 56.1
8 - 57.8
8 - 63.3
8 - 67.0
8 - 70.0
8 - 72.8

Also available in: Atom PDF