jainankitk opened a new issue, #12527: URL: https://github.com/apache/lucene/issues/12527
### Description While recently [working on numeric range queries](https://github.com/opensearch-project/OpenSearch/issues/9541), I noticed readInts24 to be consuming significant CPU cycles. When I looked into the code, I noticed [multiple consecutive invocations of readLong](https://github.com/apache/lucene/blob/main/lucene/core/src/java/org/apache/lucene/util/bkd/DocIdsWriter.java#L330). Initially, it seems that the overhead from multiple syscalls should not be as much, but I tried quick patch by reading all the longs together and it seemed to help. Sharing the patch and numbers below (nyc_taxis range query): ``` diff --git a/lucene/core/src/java/org/apache/lucene/util/bkd/DocIdsWriter.java b/lucene/core/src/java/org/apache/lucene/util/bkd/DocIdsWriter.java index 40db4c0069d..40ee7a1c968 100644 --- a/lucene/core/src/java/org/apache/lucene/util/bkd/DocIdsWriter.java +++ b/lucene/core/src/java/org/apache/lucene/util/bkd/DocIdsWriter.java @@ -325,11 +325,14 @@ final class DocIdsWriter { private static void readInts24(IndexInput in, int count, IntersectVisitor visitor) throws IOException { + long[] scratchLong = new long[(count/8) * 3]; + in.readLongs(scratchLong, 0, (count/8) * 3); int i; for (i = 0; i < count - 7; i += 8) { - long l1 = in.readLong(); - long l2 = in.readLong(); - long l3 = in.readLong(); + int li = (i/8) * 3; + long l1 = scratchLong[li]; + long l2 = scratchLong[li+1]; + long l3 = scratchLong[li+2]; visitor.visit((int) (l1 >>> 40)); visitor.visit((int) (l1 >>> 16) & 0xffffff); visitor.visit((int) (((l1 & 0xffff) << 8) | (l2 >>> 56))); ``` Without this change: ``` | Max Throughput | range | 0.71 | ops/s | | 50th percentile latency | range | 245.533 | ms | | 90th percentile latency | range | 248.005 | ms | | 99th percentile latency | range | 254.824 | ms | | 100th percentile latency | range | 256.902 | ms | | 50th percentile service time | range | 243.585 | ms | | 90th percentile service time | range | 246.178 | ms | | 99th percentile service time | range | 252.672 | ms | | 100th percentile service time | range | 255.072 | ms | | error rate | range | 0 | % | ``` With this change: ``` | Median Throughput | range | 0.7 | ops/s | | Max Throughput | range | 0.71 | ops/s | | 50th percentile latency | range | 207.554 | ms | | 90th percentile latency | range | 209.392 | ms | | 99th percentile latency | range | 213.157 | ms | | 100th percentile latency | range | 219.398 | ms | | 50th percentile service time | range | 205.421 | ms | | 90th percentile service time | range | 207.361 | ms | | 99th percentile service time | range | 211.164 | ms | | 100th percentile service time | range | 217.787 | ms | | error rate | range | 0 | % | ``` -- 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: issues-unsubscr...@lucene.apache.org.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
