NSAmelchev opened a new pull request, #13843:
URL: https://github.com/apache/ignite/pull/13843
### Problem
IGNITE-29158 replaced the class check in `BinaryUtils#isBinaryArray` (`val
instanceof BinaryArray`)
with an interface check (`val instanceof BinaryObjectEx && isArray()`).
The method is called from `CacheObjectUtils#unwrapBinary` for every column
value of every row read
by the Calcite engine (`CacheTableDescriptorImpl.FieldDescriptor#value`).
The values there are plain
`Integer`/`String` objects, so the check always fails, and on JDK 17 a
failed `instanceof` against
an interface is a linear scan over the interfaces of the value class
(JDK-8180450).
The benchmark stand reported for the IGNITE-29158 commit on the
`sql-calcite-query` scenario
(thin client, 16 threads, 1 server): throughput -5.5%, CPU per operation
+5.5%, p99 +6.7%.
### Fix
`BinaryUtils#isBinaryArray` checks the class again:
`binariesFactory.binaryArrayClass().isInstance(val)`. `binariesFactory` is a
`static final` field,
so the JIT folds the call into the same class check as before IGNITE-29158.
`BinaryObjectEx#isArray()` and its override in `BinaryArray` had no other
usages and are removed.
### Measurements
JMH, average time per value, 1 thread, values `Integer`, `String`, `Integer`.
x86_64 (AMD Ryzen 9 9950X3D, KVM, 2 vCPU, Ubuntu 26.04, OpenJDK 17.0.21):
| Method | Before IGNITE-29158 | master
| This PR |
|-------------------------------------------|---------------------|---------------|---------------|
| `CacheObjectUtils#unwrapBinary` | 10.28 ± 0.12 ns | 17.89 ±
0.12 ns | 10.23 ± 0.08 ns |
| `CacheObjectUtils#unwrapBinaryIfNeeded` | not measured | 26.47 ±
0.17 ns | 18.20 ± 0.14 ns |
aarch64 (Apple Silicon, OpenJDK 17.0.17):
| Method | Before IGNITE-29158 | master |
This PR |
|---------------------------------|---------------------|----------------|----------------|
| `CacheObjectUtils#unwrapBinary` | 3.03 ± 0.01 ns | 4.52 ± 0.04 ns |
3.10 ± 0.07 ns |
Query mix of the `sql-calcite-query` scenario (index range scan, GROUP BY,
colocated join; 100k rows)
on one embedded node, x86_64, 1 thread, 4 interleaved rounds: 3037 ops/s on
master, 3137 ops/s with
this PR (+3%, faster in 3 rounds of 4). Fork-to-fork variance on that host
is 3-5%, so this number
only agrees in direction with the stand; it was not rerun on the stand
itself.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]