[SPARK-3570] Shuffle write time does not include time to open shuffle files

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.9.2, 1.0.2, 1.1.0
• Fix Version/s: 1.3.1, 1.4.0
• Component/s: Spark Core
• Target Version/s: 1.3.1, 1.4.0
• Assignee: Kay Ousterhout
• Reporter: Kay Ousterhout
• Created: 17/Sep/14 19:57
• Updated: 24/Mar/15 23:29
• Resolved: 24/Mar/15 23:29

Description

Currently, the reported shuffle write time does not include time to open the shuffle files. This time can be very significant when the disk is highly utilized and many shuffle files exist on the machine (I'm not sure how severe this is in 1.0 onward – since shuffle files are automatically deleted, this may be less of an issue because there are fewer old files sitting around). In experiments I did, in extreme cases, adding the time to open files can increase the shuffle write time from 5ms (of a 2 second task) to 1 second. We should fix this for better performance debugging.

Thanks shivaram for helping to diagnose this problem. cc pwendell

Issue Links

• links to: [Github] Pull Request #4550 (kayousterhout)