[
https://issues.apache.org/jira/browse/COUCHDB-495?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12751215#action_12751215
]
Damien Katz commented on COUCHDB-495:
-------------------------------------
I've attach the file I'm using for performance benchmarking couch_perf.py. In
my tests, the majority of the time was spent inside the "less" comparator
function. Part of the problem is the expense of the callout the ICU for
collation, which is copying the strings to buffers before comparing them. That
can be fixed by using a more efficient method of sending data to Erlang native
ports, which is something I'm working on.
However, our json comparison function is also far more expensive than the
built-in Erlang term comparison operators.
So the "easy" solution is to just do a native Erlang term collation option.
This is the option used by views that don't need collation, just performance.
A better, but not really possible yet solution, is to code our own comparison
function in C, to be on par with Erlangs built-in comparison. I think this
isn't yet possible with Erlang and C code without hacking the core VM.
> Make views twice as fast
> ------------------------
>
> Key: COUCHDB-495
> URL: https://issues.apache.org/jira/browse/COUCHDB-495
> Project: CouchDB
> Issue Type: Improvement
> Components: JavaScript View Server
> Reporter: Chris Anderson
> Fix For: 0.11
>
> Attachments: binary_collate.diff, couch_perf.py, term_collate.diff
>
>
> Devs,
> Damien's identified view collation as the most significant bottleneck for the
> view generation. We've done some testing, and some preliminary patches, and
> the upshot seems to be that even removing ICU from the collator is not a
> significant boost. What does speed things up greatly is using raw Erlang term
> comparison. Eg, instead of using couch_view:less_json, using fun(A,B) A < B
> end. provides a roughly 2x speedup.
> However, the patch is challenging for a few reasons: Making the collation
> strategy switchable at all is tough. It's actually quite easy to get an
> alternate less function into the btree writer (all you've got to do is set it
> in couch_view_group:init_group). The hard part is propagating the same less
> function to the PassedEndFun. There's a secondary problem that when you use
> raw term comparison, a lot of terms turn out to come before nil, and after
> {}, which we use as artificial first and last terms in the less_json
> function. So just switching to raw collation alone will leave you with a view
> with unreachable rows.
> I tried two different approaches to the problem last night, and both of them
> led to (instructive) dead ends. I'll attach them for illustration purposes.
> The next line of attack we think should be tried is this:
> First - remove _all_docs_by_seq, as it is just adding complexity to the
> problem, and has been deprecated by _changes anyway. Along the same lines,
> _all_docs should no longer use couch_httpd_view:make_view_fold_fun as it has
> completely different collation needs than make_view_fold_fun. We'll end up
> duplicating a little code in the _all_docs implementation, but it should be
> worth it because it will make the other work much simpler.
> Once those changes have laid the groundwork, the next step is to change
> make_view_fold_fun and couch_view:fold, so that rather than
> make_view_fold_fun being responsible for detecting when we've passed the
> endkey. That means make_passed_end_fun and all references to PassedEnd and
> PassedEnd fun will be stripped from couch_httpd_view and moved to couch_btree.
> couch_view:fold (and the underlying btree) will need to accept not just a
> start, but also an endkey. This will make it much easier to use the less fun
> that is stored on View#view.btree#btree.less to determine PassedEnd funs.
> This will move some complexity to the btree code from the view code, but will
> keep the concerns more aligned. This also means that the btree will need to
> accept not only an endkey for folds, but also an inclusive_end parameter.
> Once we have all these refactorings done, it will be easy to make the less
> fun for an index configurable, as both the index writer and the index reader
> will look for it in the same place (on the #btree record).
> My aim is to start a discussion and get someone excited to work on this
> patch. Think of all the fast-views glory you'll get! Please ask questions and
> otherwise force me to clarify the above discussion.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
