[
https://issues.apache.org/jira/browse/HBASE-20866?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16538962#comment-16538962
]
Andrew Purtell commented on HBASE-20866:
----------------------------------------
[~reidchan] the problem is an accumulation of multiple changes over time in the
set of all changes introduced from 0.98 through to 1.3. Many thanks to
[~vik.karma] for finding this one. There will be more, I am sure. As Vikas said
{quote}
To identify the root cause of performance diff between 0.98 and 1.3 we carried
out lot of experiments with profiling and git bisect iterations, however we
were not able to identify any particular source of scan performance degradation
and it looked like this is an accumulated degradation of 5-10% over various
enhancements and refactoring.
{quote}
I'm raising the priority to Critical because this impacts anyone seriously
considering a migration off of 0.98 to a 1.x, and anyone who already has.
> HBase 1.x scan performance degradation compared to 0.98 version
> ---------------------------------------------------------------
>
> Key: HBASE-20866
> URL: https://issues.apache.org/jira/browse/HBASE-20866
> Project: HBase
> Issue Type: Bug
> Affects Versions: 1.3.2
> Reporter: Vikas Vishwakarma
> Priority: Major
> Fix For: 1.5.0, 1.2.7, 1.3.3, 1.4.6
>
>
> Internally while testing 1.3 as part of migration from 0.98 to 1.3 we
> observed perf degradation in scan performance for phoenix queries varying
> from few 10's to upto 200% depending on the query being executed. We tried
> simple native HBase scan and there also we saw upto 40% degradation in
> performance when the number of column qualifiers are high (40-50+)
> To identify the root cause of performance diff between 0.98 and 1.3 we
> carried out lot of experiments with profiling and git bisect iterations,
> however we were not able to identify any particular source of scan
> performance degradation and it looked like this is an accumulated degradation
> of 5-10% over various enhancements and refactoring.
> We identified few major enhancements like partialResult handling,
> ScannerContext with heartbeat processing, time/size limiting, RPC
> refactoring, etc that could have contributed to small degradation in
> performance which put together could be leading to large overall degradation.
> One of the changes is
> [HBASE-11544|https://jira.apache.org/jira/browse/HBASE-11544] which
> implements partialResult handling. In ClientScanner.java the results received
> from server are cached on the client side by converting the result array into
> an ArrayList. This function gets called in a loop depending on the number of
> rows in the scan result. Example for ten’s of millions of rows scanned, this
> can be called in the order of millions of times.
> In almost all the cases 99% of the time (except for handling partial results,
> etc). We are just taking the resultsFromServer converting it into a ArrayList
> resultsToAddToCache in addResultsToList(..) and then iterating over the list
> again and adding it to cache in loadCache(..) as given in the code path below
> In ClientScanner → loadCache(..) → getResultsToAddToCache(..) →
> addResultsToList(..) →
> {code:java}
> loadCache() {
> ...
> List<Result> resultsToAddToCache =
> getResultsToAddToCache(values, callable.isHeartbeatMessage());
> ...
> …
> for (Result rs : resultsToAddToCache) {
> rs = filterLoadedCell(rs);
> cache.add(rs);
> ...
> }
> }
> getResultsToAddToCache(..) {
> ..
> final boolean isBatchSet = scan != null && scan.getBatch() > 0;
> final boolean allowPartials = scan != null &&
> scan.getAllowPartialResults();
> ..
> if (allowPartials || isBatchSet) {
> addResultsToList(resultsToAddToCache, resultsFromServer, 0,
> (null == resultsFromServer ? 0 : resultsFromServer.length));
> return resultsToAddToCache;
> }
> ...
> }
> private void addResultsToList(List<Result> outputList, Result[] inputArray,
> int start, int end) {
> if (inputArray == null || start < 0 || end > inputArray.length) return;
> for (int i = start; i < end; i++) {
> outputList.add(inputArray[i]);
> }
> }{code}
>
> It looks like we can avoid the result array to arraylist conversion
> (resultsFromServer --> resultsToAddToCache ) for the first case which is also
> the most frequent case and instead directly take the values arraay returned
> by callable and add it to the cache without converting it into ArrayList.
> I have taken both these flags allowPartials and isBatchSet out in loadcahe()
> and I am directly adding values to scanner cache if the above condition is
> pass instead of coverting it into arrayList by calling
> getResultsToAddToCache(). For example:
> {code:java}
> protected void loadCache() throws IOException {
> Result[] values = null;
> ..
> final boolean isBatchSet = scan != null && scan.getBatch() > 0;
> final boolean allowPartials = scan != null && scan.getAllowPartialResults();
> ..
> for (;;) {
> try {
> values = call(callable, caller, scannerTimeout);
> ..
> } catch (DoNotRetryIOException | NeedUnmanagedConnectionException e) {
> ..
> }
> if (allowPartials || isBatchSet) { // DIRECTLY COPY values TO CACHE
> if (values != null) {
> for (int v=0; v<values.length; v++) {
> Result rs = values[v];
> ....
> cache.add(rs);
> ...
> } else { // DO ALL THE REGULAR PARTIAL RESULT HANDLING ..
> List<Result> resultsToAddToCache =
> getResultsToAddToCache(values, callable.isHeartbeatMessage());
> for (Result rs : resultsToAddToCache) {
> ....
> cache.add(rs);
> ...
> }
> }
> {code}
>
> I am seeing upto 10% improvement in scan time with these changes, sample PE
> execution results given below.
> ||PE (1M , 1 thread)||with addResultsToList||without
> addResultsToList||%improvement||
> |ScanTest|9228|8448|9|
> |RandomScanWithRange10Test|393413|378222|4|
> |RandomScanWithRange100Test|1041860|980147|6|
> Similarly we are observing upto 10% improvement in simple native HBase scan
> test used internally that just scans through a large region filtering all the
> rows. I still have to do the phoenix query tests with this change. Posting
> the initial observations for feedback/comments and suggestions.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
