IO

• Background
  • Seek vs Read
  • ORC Read
• Read Optimization
  • Approach
  • Scope
  • Benchmarks
    • Local FS
    • AWS S3
  • Summary

Background

We are moving our workloads from HDFS to AWS S3. As part of this activity we wanted to understand the performance characteristics and costs of using S3.

Seek vs Read

One particular scenario that stood out in our performance testing was Seek vs Read when dealing with S3.

In this test we are trying to read through a file

• Seek to Point A in the file read X bytes
• Move to Point B in the file that is A + X + Y
  • This is accomplished as another seek or as a read
  • We will leave Y variable to determine when this is best
• Read X bytes

Observations:

• We could clearly see that a read is more performant than seek when dealing with steps/gaps smaller than 4 MB.
  • At 4 MB read is faster by ~ 11%
  • At 1 MB read is faster by ~ 20%
• Reads are also cheaper as we perform a single GET instead of multiple GETs from AWS S3 Pricing
  • Cost for GET: $0.0004
  • Cost for Data Retrieval to the same region AWS EKS: $0.0000

ORC Read

Based on the above performance penalty when dealing with multiple seeks over small gaps, we measured the performance of ORC read on a file.

File details:

• Size ~ 21 MB
• Column Count: ~ 400
• Row Count: ~ 65K

Read Type	Duration	Unit
All Columns	1.075	s
Alternate Columns	6.489	s

Observations:

• We can clearly see that we pay a significant penalty when reading alternate columns, which in the current implementation of ORC translates to multiple GET calls on AWS S3
• While the impact of penalty will be less significant in large reads, it will incur overheads both in terms of time and cost

Read Optimization

Approach

The following optimizations are planned:

• orc.min.disk.seek.size is a value in bytes: When trying to determine a single read, if the gap between two reads is smaller than this then it is combined into a single read.
• orc.min.disk.seek.size.tolerance is a fractional input: If the extra bytes read is greater than this fraction of the required bytes, then we drop the extra bytes from memory.
• We can further consider adding an optimization for the complete stripe in case the stripe size is smaller than orc.min.disk.seek.size

Scope

Different types of IO takes place in ORC today.

• Reading of File Footer: Unchanged
• Reading of Stripe Footer: Unchanged
• Reading of Stripe Index information: Optimized
• Reading of Stripe Data: Optimized

Each of the above happens at different stages of the read. The current implementation optimizes reads that happen using the DataReader interface.

This does not:

• Optimize the read of the file/stripe footer
• Reads across multiple stripes

Benchmarks

Local FS

This benchmark is run on the local filesystem with NVMe SSD, so it has very different performance characteristics to AWS S3.

The purpose of this benchmark is to ascertain if we have added any significant penalties in the ORC code by adding minSeekSize and extraByteTolerance.

java -jar java/bench/core/target/orc-benchmarks-core-*-uber.jar chunk_read

(alt)	(cols)	(byteTol)	(minSeek)	Mode	Cnt	Score	Sign	Error	Units
true	128	0.0	0	avgt	20	0.352	±	0.006	s/op
true	128	0.0	4194304	avgt	20	0.357	±	0.002	s/op
true	128	10.0	4194304	avgt	20	0.349	±	0.002	s/op
false	128	0.0	0	avgt	20	0.667	±	0.007	s/op
false	128	0.0	4194304	avgt	20	0.673	±	0.004	s/op
false	128	10.0	4194304	avgt	20	0.671	±	0.005	s/op

Observations/Details:

• Input File details:
  • Rows: 65536
  • Columns: 128
  • FileSize: ~ 72 MB
• Full Read (alternate = false)
  • No significant difference between the options as expected
• Alternate Read (alternate = true)
  • No significant difference between the options given the small file size and performance of local disk
  • This also calls out that the recommended minSeekSize should be determined for each platform e.g. HDFS, S3, etc

AWS S3

In this benchmark we brought up an EKS Container in the same region as the AWS S3 bucket to test the performance of the patch.

(alternate)	(byteTol)	(minSeekSize)	Mode	Cnt	Score	Sign	Error	Units
FALSE	0.0	0	avgt	5	1.837	±	0.089	s/op
FALSE	0.0	4194304	avgt	5	1.919	±	0.11	s/op
FALSE	10.0	4194304	avgt	5	1.895	±	0.191	s/op
TRUE	0.0	0	avgt	5	5.8	±	1.132	s/op
TRUE	0.0	4194304	avgt	5	1.479	±	0.197	s/op
TRUE	10.0	4194304	avgt	5	1.435	±	0.176	s/op

Observations/Details:

• Input File details:
  • Rows: 65536
  • Columns: 128
  • FileSize: ~ 72 MB
• Full Read (alternate = false)
  • No significant difference between the options as expected
• Alternate Read (alternate = true)
  • We get a significant boost in performance 5.8s without optimization to 1.5s with optimization giving us a time reduction of ~ 75 %
  • This also gives us a cost saving as 64 GET one for each column per stripe have been replaced with a single GET
  • We can see a marginal improvement ~ 3% when choosing to retain extra bytes (extraByteTolerance=10.0) as compared to (extraByteTolerance=0.0) which performs additional work of dropping the extra bytes from memory.

Summary

Based on the benchmarks the following is recommended for ORC in AWS S3:

• orc.min.disk.seek.size is set to 4194304 (4 MB)
• orc.min.disk.seek.size.tolerance is set to value that is acceptable based on the memory usage constraints. When set to 0.0 it will always do the extra work of dropping the extra bytes.