[
https://issues.apache.org/jira/browse/LUCENE-9850?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17315066#comment-17315066
]
Greg Miller commented on LUCENE-9850:
-------------------------------------
Thanks [~mikemccand]! For starters, yes—all the runs I referenced here are
using "wikimediumall".
{quote}I would expect {{XTerm}} to show speedups since this is largely
dominated by decoding many postings blocks. But it is odd to see the
{{XTermYSort}} tasks negatively impacted: those tasks are just sorting by a
{{DocValues}} field instead of default text relevance (BM25).
{quote}
I might except the opposite actually. Anytime the PFOR approach has to apply
exceptions, I would expect a performance hit of some sort since it has extra
work to do on top of the FOR approach used today. So if the Term tasks are
largely dominated by postings decoding, I would expect regressions to show up
there more than elsewhere. Maybe I'm misunderstanding your comment though?
I re-ran wikimediumall with only "Term" tasks and got the following (looks like
noise to me):
{code:java}
TaskQPS baseline StdDevQPS pfordocids StdDev
Pct diff p-value
HighTermDayOfYearSort 5.74 (10.8%) 5.59 (9.9%)
-2.6% ( -20% - 20%) 0.431
TermDTSort 44.54 (15.4%) 44.06 (14.3%)
-1.1% ( -26% - 33%) 0.816
HighTermTitleBDVSort 30.90 (14.3%) 30.59 (13.5%)
-1.0% ( -25% - 31%) 0.820
MedTerm 392.89 (6.8%) 389.45 (7.5%)
-0.9% ( -14% - 14%) 0.699
LowTerm 412.80 (7.0%) 410.68 (7.9%)
-0.5% ( -14% - 15%) 0.827
PKLookup 130.70 (2.8%) 131.44 (2.0%)
0.6% ( -4% - 5%) 0.470
HighTermMonthSort 61.69 (12.2%) 62.13 (13.6%)
0.7% ( -22% - 30%) 0.860
HighTerm 381.73 (10.0%) 385.03 (7.8%)
0.9% ( -15% - 20%) 0.761
{code}
I also pulled out all the tasks in my last wikimediumall run that had a
significant change in either direction (p-value <= 0.05) and reran them alone
to see if the results were repeatable. In general, they were not. The three
that *were* repeatable (significant) regressions were LowSpanNear (-2.2%),
AndHighMed (-2.1%) and AndHighHigh (-2.0%):
{code:java}
TaskQPS baseline StdDevQPS pfordocids StdDev
Pct diff p-value
HighTermTitleBDVSort 42.09 (10.6%) 41.04 (10.4%)
-2.5% ( -21% - 20%) 0.451
LowSpanNear 4.29 (1.9%) 4.20 (1.4%)
-2.2% ( -5% - 1%) 0.000
AndHighMed 26.79 (3.0%) 26.23 (2.4%)
-2.1% ( -7% - 3%) 0.014
AndHighHigh 13.83 (3.4%) 13.54 (2.8%)
-2.0% ( -7% - 4%) 0.037
HighTerm 696.28 (6.7%) 688.02 (7.0%)
-1.2% ( -13% - 13%) 0.585
OrNotHighHigh 372.40 (5.3%) 370.23 (5.3%)
-0.6% ( -10% - 10%) 0.726
HighTermDayOfYearSort 13.53 (9.8%) 13.46 (10.4%)
-0.5% ( -18% - 21%) 0.866
MedSpanNear 19.76 (1.7%) 19.67 (1.4%)
-0.5% ( -3% - 2%) 0.320
HighTermMonthSort 48.94 (11.2%) 48.77 (10.5%)
-0.3% ( -19% - 24%) 0.921
LowTerm 714.77 (3.0%) 714.27 (5.6%)
-0.1% ( -8% - 8%) 0.960
PKLookup 139.66 (3.0%) 139.69 (2.4%)
0.0% ( -5% - 5%) 0.984
IntNRQ 33.49 (1.4%) 33.54 (1.2%)
0.2% ( -2% - 2%) 0.690
{code}
Finally, I tried running a micro-benchmark to see if I could isolate how much
regression there was due to applying the exceptions in PFOR compared to FOR. I
forked [~jpountz]'s code originally used for optimizing FOR. The results are in
the README over
[here|https://github.com/gsmiller/decode-128-ints-benchmark/tree/pfor-delta] if
interested. They generally make sense to me and show that performance does take
a hit when there are exceptions to patch in, but I think we're seeing with the
overall luceneutil benchmark's that those performance shifts aren't showing up
in the bigger picture for the most part.
So overall, it seems like the index size reduction might be worth it here, but
I'm new to Lucene benchmarks and this is my first attempt at running a
micro-benchmark, so I would trust the opinion of others with more experienced
eyes here.
> Explore PFOR for Doc ID delta encoding (instead of FOR)
> -------------------------------------------------------
>
> Key: LUCENE-9850
> URL: https://issues.apache.org/jira/browse/LUCENE-9850
> Project: Lucene - Core
> Issue Type: Task
> Components: core/codecs
> Affects Versions: main (9.0)
> Reporter: Greg Miller
> Priority: Minor
> Attachments: apply_exceptions.png, bulk_read_1.png, bulk_read_2.png,
> for.png, pfor.png
>
>
> It'd be interesting to explore using PFOR instead of FOR for doc ID encoding.
> Right now PFOR is used for positions, frequencies and payloads, but FOR is
> used for doc ID deltas. From a recent
> [conversation|http://mail-archives.apache.org/mod_mbox/lucene-dev/202103.mbox/%3CCAPsWd%2BOp7d_GxNosB5r%3DQMPA-v0SteHWjXUmG3gwQot4gkubWw%40mail.gmail.com%3E]
> on the dev mailing list, it sounds like this decision was made based on the
> optimization possible when expanding the deltas.
> I'd be interesting in measuring the index size reduction possible with
> switching to PFOR compared to the performance reduction we might see by no
> longer being able to apply the deltas in as optimal a way.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
