Pavan Lanka created ORC-1136:
--------------------------------
Summary: Optimize reads by combining multiple reads without
significant separation into a single read
Key: ORC-1136
URL: https://issues.apache.org/jira/browse/ORC-1136
Project: ORC
Issue Type: Improvement
Components: Java
Affects Versions: 1.7.3
Reporter: Pavan Lanka
Assignee: Pavan Lanka
Fix For: 1.8.0
h2. 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.
h3. 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
!seekvsread.png!
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|https://aws.amazon.com/s3/pricing/]
** Cost for GET: $0.0004
** Cost for Data Retrieval to the same region AWS EKS: $0.0000
h3. 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
!image-2022-03-25-16-45-00-692.png!
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
h2. Read Optimization
The following optimizations are proposed:
* *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`
h3. 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
{color:#287bde}DataReader{color}{color:#cc7832} {color}interface.
This does not:
* Optimize the read of the file/stripe footer
* Reads across multiple stripes
h2. Benchmarks
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.
!benchmark.png!
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.
h2. 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.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
