Another interesting thing...during my profiling activity I gave a chance to the old v2 protocol and activated the gc logs, as expected the result is that with v2 protocol there is almost no GC activity during the benchmark -- Enrico
2017-07-11 12:07 GMT+02:00 Enrico Olivelli <[email protected]>: > > > 2017-07-11 11:04 GMT+02:00 Sijie Guo <[email protected]>: > >> I think Netty4 requires more offheap memory. you might need to tune the >> JVM >> settings. I doubt that latency diff coming from the JVM gc. >> >> A simple thing to verify that is to dump the gc log by adding " -Xloggc:" >> setting and compare the gc logs between versions. >> > > > Enabling G1 GC (-XX:+UseG1GC) apparently solves the difference between 4.4 > and 4.5. > > And I a loop the bench (in order to have a warm JVM) numbers become more > stable and similar to each other > > I have the "other issue" pending, the fact that latency (time for > asynchAddEntry to complete) is really bad and unpredictable (from 100 ms to > 2000ms) > > and if I introduce a Thread.sleep(1) all the callbacks complete always > with success in < 10ms. This happens even with G1 GC. > > Without the "sleep(1)" my machine uses a lot of CPU (I have 8 "CPUs") and > with the "sleep(1)" the load is slightly lower > > Honestly for me this is still a problem and I hope that with your help I > will be able to find the problem, wherever it is (in BK code or in the way > I am doing the bench) > > I will try to create a new more complete bench > > -- Enrico > > > > >> >> - Sijie >> >> On Tue, Jul 11, 2017 at 12:16 AM, Enrico Olivelli <[email protected]> >> wrote: >> >> > a did a bisect and the culprit (in my opinion) is the switch to netty 4 >> for >> > the performance regression from 4.5 and 4.4 >> > >> > at commit: >> > commit 811ece53a1c975c4e768422f3d622ac9de6b3e41 BOOKKEEPER-1058: >> Ignore >> > already deleted ledger on replication audit >> > >> > Total time: 204 ms >> > Total real time: 79 ms per entry >> > >> > at commit: >> > commit 74f795136c1fff3badb29fc982d0cc2d43096b45 BOOKKEEPER-1008: Netty >> 4.1 >> > >> > Total time: 308 ms >> > Total real time: 189 ms per entry >> > >> > I have tried with epoll and with local transport, results does not >> change. >> > I tried to upgrade to netty 4.1.13 too, but no change >> > >> > Could it be the memory allocator of netty which is overwhelmed with >> sudden >> > bursts of allocation ? >> > I did some trial with UnpooledByteBufAllocator.DEFAULT and it helps a >> > little, we get to "110 ms per entry" vs "189 ms per entry" >> > >> > the bench is here: >> > https://github.com/eolivelli/bookkeepers-benchs/blob/master/ >> src/test/java/ >> > BookKeeperWriteTest.java >> > >> > >> > -- Enrico >> > >> > >> > >> > 2017-07-10 19:46 GMT+02:00 Enrico Olivelli <[email protected]>: >> > >> > > >> > > >> > > Il lun 10 lug 2017, 18:21 Venkateswara Rao Jujjuri <[email protected] >> > >> > ha >> > > scritto: >> > > >> > >> With Netty changes, lack of native epoll() has huge perf impact as >> per >> > >> Kishore. >> > >> Are you sure you are using epoll()? >> > >> >> > > >> > > Yes. I tried with netty local transport too. It seems not related to >> > netty >> > > to me. >> > > I will double check, tomorrow >> > > Enrico >> > > >> > > >> > >> On Mon, Jul 10, 2017 at 1:49 AM, Enrico Olivelli < >> [email protected]> >> > >> wrote: >> > >> >> > >> > 2017-07-10 10:40 GMT+02:00 Sijie Guo <[email protected]>: >> > >> > >> > >> > > Also one other thing to check is the JVM settings. Do you mind >> > sharing >> > >> > that >> > >> > > as well? >> > >> > > >> > >> > > >> > >> > this is the surefire config, I am using oracle jdk 8 >> > >> > >> > >> > <plugin> >> > >> > <artifactId>maven-surefire-plugin</artifactId> >> > >> > <version>2.20</version> >> > >> > <configuration> >> > >> > <forkCount>1</forkCount> >> > >> > <reuseForks>false</reuseForks> >> > >> > >> > >> > <forkedProcessTimeoutInSeconds>300</forkedProcessTimeoutInSeconds> >> > >> > <argLine>-Xmx2G >> > >> > -Djava.io.tmpdir=${basedir}/target</argLine> >> > >> > </configuration> >> > >> > </plugin> >> > >> > >> > >> > -- Enrico >> > >> > >> > >> > >> > >> > >> > >> > > Sijie >> > >> > > >> > >> > > On Jul 10, 2017 1:17 AM, "Sijie Guo" <[email protected]> wrote: >> > >> > > >> > >> > > > I am not sure if there is any default values changed for >> journal >> > >> > > settings. >> > >> > > > I would suggest you testing by setting specifically the journal >> > >> > settings. >> > >> > > > >> > >> > > > Also if you can share your benchmark, that would be good for >> other >> > >> > people >> > >> > > > to verify. >> > >> > > > >> > >> > > > Sijie >> > >> > > > >> > >> > > > On Jul 10, 2017 12:32 AM, "Enrico Olivelli" < >> [email protected]> >> > >> > wrote: >> > >> > > > >> > >> > > >> Hi, >> > >> > > >> I am doing some benchmarks on BK, I see that from 4.4.0 to >> 4.5.0 >> > >> there >> > >> > > is >> > >> > > >> something "slow" but I cannot understand what. I really hope >> that >> > >> I am >> > >> > > >> wrong. >> > >> > > >> >> > >> > > >> I am working with writes, I will pass to reads once writes >> will >> > be >> > >> ok. >> > >> > > >> My problem is both on latency (time for AddComplete callback >> to >> > >> > > complete) >> > >> > > >> and on overall throuput. >> > >> > > >> >> > >> > > >> Actually I have two distinct problems, but working on the >> first >> > >> > problem >> > >> > > I >> > >> > > >> found a performance regression. >> > >> > > >> I know that talking about "slow" things it is an hard matter, >> so >> > I >> > >> > will >> > >> > > >> try >> > >> > > >> do describe as much as possible all the aspects that I think >> are >> > >> > > relevant. >> > >> > > >> >> > >> > > >> First problem: under certain load performance >> > (latency+throughput) >> > >> > > degrade >> > >> > > >> too much >> > >> > > >> Second problem: the first problem is more evident in 4.5.0 >> > >> > > >> >> > >> > > >> Let's describe my testcase and why I am worried. >> > >> > > >> The bench issues a batch of asyncAddEntry and prints the >> average >> > >> time >> > >> > > for >> > >> > > >> AddComplete to complete and the overall clock time. >> > >> > > >> >> > >> > > >> This is the code >> > >> > > >> >> > >> > > >> private static final byte[] TEST_DATA = new byte[35 * 1024]; >> > >> > > >> private static final int testsize = 1000; >> > >> > > >> >> > >> > > >> ...... (start 1 bookie, see below) >> > >> > > >> ClientConfiguration clientConfiguration = new >> > >> > > >> ClientConfiguration(); >> > >> > > >> clientConfiguration.setZkServ >> ers(env.getAddress()); >> > >> > > >> try (BookKeeper bk = new BookKeeper( >> > >> clientConfiguration); >> > >> > > >> LedgerHandle lh = bk.createLedger(1, 1, 1, >> > >> > > >> BookKeeper.DigestType.CRC32, new byte[0])) { >> > >> > > >> LongAdder totalTime = new LongAdder(); >> > >> > > >> long _start = System.currentTimeMillis(); >> > >> > > >> Collection<CompletableFuture> batch = new >> > >> > > >> ConcurrentLinkedQueue<>(); >> > >> > > >> for (int i = 0; i < testsize; i++) { >> > >> > > >> CompletableFuture cf = new >> > CompletableFuture(); >> > >> > > >> batch.add(cf); >> > >> > > >> lh.asyncAddEntry(TEST_DATA, new >> > >> > > >> AsyncCallback.AddCallback() { >> > >> > > >> >> > >> > > >> long start = >> System.currentTimeMillis(); >> > >> > > >> >> > >> > > >> @Override >> > >> > > >> public void addComplete(int rc, >> > >> LedgerHandle >> > >> > lh, >> > >> > > >> long entryId, Object ctx) { >> > >> > > >> long now = >> > >> > > >> System.currentTimeMillis(); >> > >> > > >> CompletableFuture _cf = >> > >> > (CompletableFuture) >> > >> > > >> ctx; >> > >> > > >> if (rc == BKException.Code.OK) { >> > >> > > >> _cf.complete(""); >> > >> > > >> } else { >> > >> > > >> >> > >> > > >> _cf.completeExceptionally(BKException.create(rc)); >> > >> > > >> } >> > >> > > >> totalTime.add(now - start); >> > >> > > >> } >> > >> > > >> }, cf); >> > >> > > >> // Thread.sleep(1); // this is the >> > >> tirgger!!! >> > >> > > >> } >> > >> > > >> assertEquals(testsize, batch.size()); >> > >> > > >> for (CompletableFuture f : batch) { >> > >> > > >> f.get(); >> > >> > > >> } >> > >> > > >> long _stop = System.currentTimeMillis(); >> > >> > > >> long delta = _stop - _start; >> > >> > > >> System.out.println("Total time: " + delta + " >> > ms"); >> > >> > > >> System.out.println("Total real time: " + >> > >> > > totalTime.sum() + >> > >> > > >> " ms -> "+(totalTime.sum()/testsize)+" ms per entry"); >> > >> > > >> } >> > >> > > >> >> > >> > > >> Bookie config: >> > >> > > >> ServerConfiguration conf = new ServerConfiguration(); >> > >> > > >> conf.setBookiePort(5621); >> > >> > > >> conf.setUseHostNameAsBookieID(true); >> > >> > > >> >> > >> > > >> Path targetDir = path.resolve("bookie_data"); >> > >> > > >> conf.setZkServers("localhost:1282"); >> > >> > > >> conf.setLedgerDirNames(new >> > >> > > >> String[]{targetDir.toAbsolutePath().toString()}); >> > >> > > >> conf.setJournalDirName(targetDir.toAbsolutePath(). >> > >> > toString()); >> > >> > > >> conf.setFlushInterval(1000); >> > >> > > >> conf.setJournalFlushWhenQueueEmpty(true); >> > >> > > >> conf.setProperty("journalMaxGroupWaitMSec", 0); >> > >> > > >> conf.setProperty("journalBufferedWritesThreshold", >> > 1024); >> > >> > > >> conf.setAutoRecoveryDaemonEnabled(false); >> > >> > > >> conf.setEnableLocalTransport(true); >> > >> > > >> conf.setAllowLoopback(true); >> > >> > > >> >> > >> > > >> The tests starts one ZK server + 1 Bookie + the testcase in a >> > JUnit >> > >> > test >> > >> > > >> >> > >> > > >> >> > >> > > >> Results: >> > >> > > >> A - BK-4.4.0: >> > >> > > >> Total time: 209 ms >> > >> > > >> Total real time: 194337 ms -> 194 ms per entry >> > >> > > >> >> > >> > > >> B - BK-4.5.0-SNAPSHOT: >> > >> > > >> Total time: 269 ms >> > >> > > >> Total real time: 239918 ms -> 239 ms per entry >> > >> > > >> >> > >> > > >> C - BK-4.4,0 with sleep(1): >> > >> > > >> Total time: 1113 ms (1000 ms sleep time) >> > >> > > >> Total real time: 4238 ms -> 4 ms per entry >> > >> > > >> >> > >> > > >> D - BK-4.5,0-SNAPSHOT with sleep(1): >> > >> > > >> Total time: 1121 ms (1000 ms sleep time) >> > >> > > >> Total real time: 8018 ms -> 8 ms per entry >> > >> > > >> >> > >> > > >> Problem 1 (unexpected performance degradation): >> > >> > > >> Times per entry (latency) are incredibly slow in cases A and >> B. >> > >> > > >> If I add a sleep(1) between one call of asyncAddEntry and the >> > next >> > >> > > >> "latency" is around 4 ms per entry. >> > >> > > >> >> > >> > > >> Problem 2: worse performance on 4.5.0 >> > >> > > >> Compare A vs B and C vs D, it is self-explaining. >> > >> > > >> >> > >> > > >> I am running the test on my laptop, with linux 64bit >> (Fedora), 12 >> > >> GB >> > >> > > RAM, >> > >> > > >> no swap, on an SSD disk. The results are similar on other >> > >> computers. >> > >> > > >> >> > >> > > >> It seems that if I issue too many addEntry the systems slows >> > down. >> > >> > > >> >> > >> > > >> Please note this fact: >> > >> > > >> numbers for case A and B (without sleep) mean that all the >> adds >> > got >> > >> > > >> completed almost together >> > >> > > >> >> > >> > > >> for the 4.5 vs 4.4 case: >> > >> > > >> I tried to disable all of the threadpool enhancements >> (different >> > >> > > >> read/write >> > >> > > >> pools)....it makes not difference >> > >> > > >> >> > >> > > >> Questions: >> > >> > > >> >> > >> > > >> Is the "grouping" logic of the journal ? >> > >> > > >> >> > >> > > >> Is there a way of making a burst of 1000 async writes on the >> same >> > >> > ledger >> > >> > > >> perform <10 ms latency ? (in my real case I have bursts of >> > >> concurrent >> > >> > > >> writes from different threads) >> > >> > > >> >> > >> > > >> Why 4.5.0 is anyway slower ? >> > >> > > >> >> > >> > > >> Thanks >> > >> > > >> >> > >> > > >> -- Enrico >> > >> > > >> >> > >> > > > >> > >> > > >> > >> > >> > >> >> > >> >> > >> >> > >> -- >> > >> Jvrao >> > >> --- >> > >> First they ignore you, then they laugh at you, then they fight you, >> then >> > >> you win. - Mahatma Gandhi >> > >> >> > > -- >> > > >> > > >> > > -- Enrico Olivelli >> > > >> > >> > >
