[
https://issues.apache.org/jira/browse/LUCENE-10572?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17537410#comment-17537410
]
Dawid Weiss commented on LUCENE-10572:
--------------------------------------
> Nevertheless, the main limiting factor of the BytesRefHash is the equals
> (although vectorized) because it always needs to be verified
Right. This strikes the nostalgic note of the strlen performance between pascal
and C, doesn't it?... :) This is such a hot code section that indeed storing
the length along the string itself may be worth it. I still use that "offset
difference" strategy in non-Lucene code where it performs quite well but it's
really a matter of trying and I bet the results will vary depending on the
context (terms, caches, etc.).
> we can lookup offset of next entry - offset of entry to be looked up. The
> only special case is the very last item.
This can be solved elegantly and efficiently - the offsets array stores the
end+1 of each element, with the initial 0-offset index initially set to zero.
So, the length of entry i is a constant expression (offsets[i + 1] -
offsets[i]) and this invariant is maintained upon additions of new elements
like so:
bytePool.add(ref.bytes, ref.offset, ref.length);
offsets.add(bytePool.size());
This invariant makes all the remaining functions simpler too, for example
element-comparing method is something like this (code copy-pasted from ours,
but you'll get the gist):
{code}
public int compare(int elementA, int elementB) {
assert elementA >= 0 && elementA < size() && elementB >= 0 && elementB <
size();
int off1 = offsets.get(elementA);
int len1 = offsets.get(elementA + 1) - off1;
int off2 = offsets.get(elementB);
int len2 = offsets.get(elementB + 1) - off2;
return Bytes.compare(blocks.buffer, off1, len1, blocks.buffer, off2, len2);
}
{code}
The caveat here is that the offsets array is an int[] so the storage size
required for the hashes is slightly higher. Overall this was never a problem in
practice though.
> Can we optimize BytesRefHash?
> -----------------------------
>
> Key: LUCENE-10572
> URL: https://issues.apache.org/jira/browse/LUCENE-10572
> Project: Lucene - Core
> Issue Type: Improvement
> Reporter: Michael McCandless
> Priority: Major
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> I was poking around in our nightly benchmarks
> ([https://home.apache.org/~mikemccand/lucenebench]) and noticed in the JFR
> profiling that the hottest method is this:
> {noformat}
> PERCENT CPU SAMPLES STACK
> 9.28% 53848 org.apache.lucene.util.BytesRefHash#equals()
> at
> org.apache.lucene.util.BytesRefHash#findHash()
> at org.apache.lucene.util.BytesRefHash#add()
> at
> org.apache.lucene.index.TermsHashPerField#add()
> at
> org.apache.lucene.index.IndexingChain$PerField#invert()
> at
> org.apache.lucene.index.IndexingChain#processField()
> at
> org.apache.lucene.index.IndexingChain#processDocument()
> at
> org.apache.lucene.index.DocumentsWriterPerThread#updateDocuments() {noformat}
> This is kinda crazy – comparing if the term to be inserted into the inverted
> index hash equals the term already added to {{BytesRefHash}} is the hottest
> method during nightly benchmarks.
> Discussing offline with [~rcmuir] and [~jpountz] they noticed a few
> questionable things about our current implementation:
> * Why are we using a 1 or 2 byte {{vInt}} to encode the length of the
> inserted term into the hash? Let's just use two bytes always, since IW
> limits term length to 32 K (< 64K that an unsigned short can cover)
> * Why are we doing byte swapping in this deep hotspot using {{VarHandles}}
> (BitUtil.VH_BE_SHORT.get)
> * Is it possible our growth strategy for {{BytesRefHash}} (on rehash) is not
> aggressive enough? Or the initial sizing of the hash is too small?
> * Maybe {{MurmurHash}} is not great (causing too many conflicts, and too
> many {{equals}} calls as a result?) – {{Fnv}} and {{xxhash}} are possible
> "upgrades"?
> * If we stick with {{{}MurmurHash{}}}, why are we using the 32 bit version
> ({{{}murmurhash3_x86_32{}}})?
> * Are we using the JVM's intrinsics to compare multiple bytes in a single
> SIMD instruction ([~rcmuir] is quite sure we are indeed)?
> * [~jpountz] suggested maybe the hash insert is simply memory bound
> * {{TermsHashPerField.writeByte}} is also depressingly slow (~5% of total
> CPU cost)
> I pulled these observations from a recent (5/6/22) profiler output:
> [https://home.apache.org/~mikemccand/lucenebench/2022.05.06.06.33.00.html]
> Maybe we can improve our performance on this crazy hotspot?
> Or maybe this is a "healthy" hotspot and we should leave it be!
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
