[SPARK-5269] BlockManager.dataDeserialize always creates a new serializer instance

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Won't Fix
• Component/s: Spark Core
• Labels: performance serializers
• Assignee: Unassigned
• Reporter: Ivan Vergiliev
• Created: 15/Jan/15 15:33
• Updated: 16/Dec/15 10:56
• Resolved: 16/Dec/15 10:56

Description

BlockManager.dataDeserialize always creates a new instance of the serializer, which is pretty slow in some cases. I'm using Kryo serialization and have a custom registrator, and its register method is showing up as taking about 15% of the execution time in my profiles. This started happening after I increased the number of keys in a job with a shuffle phase by a factor of 40.

One solution I can think of is to create a ThreadLocal SerializerInstance for the defaultSerializer, and only create a new one if a custom serializer is passed in. AFAICT a custom serializer is passed only from DiskStore.getValues, and that, on the other hand, depends on the serializer passed to ExternalSorter. I don't know how often this is used, but I think this can still be a good solution for the standard use case.
Oh, and also - ExternalSorter already has a SerializerInstance, so if the getValues method is called from a single thread, maybe we can pass that directly?

I'd be happy to try a patch but would probably need a confirmation from someone that this approach would indeed work (or an idea for another).

Issue Links

• is related to: SPARK-3386 Reuse serializer and serializer buffer in shuffle block iterator
• links to: [Github] Pull Request #8292 (maciej-kisiel)