kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jay Kreps <...@confluent.io>
Subject Re: Performance Test Inquiries
Date Tue, 01 Sep 2015 17:15:42 GMT
Yeah okay never mind I misunderstood. I'm not sure of the cause either--I
didn't see that in my testing.

The spindle thing tends to get overestimated since our writes are async.
There is some perf hit from having multiple partitions (maybe 10-20%) but
mostly the OS does a good job of scheduling writes in order.

There is a strong effect from batching--in general fewer partitions will
lead to more efficient batching and hence more efficient network and
filesystem usage. My expectation, though, would be that you would get
plenty of batching in either test because the throughput is so high. You
could investigate this by setting linger.ms=1 and see if that changes
anything.

-Jay

On Mon, Aug 31, 2015 at 8:14 PM, explorer <jindotw@gmail.com> wrote:

> Hi Jay,
>
> For issue #1, I will file a JIRA so the community or dev team can take
> a look at it.
>
> For my 2nd questions, the scenario goes like this:
>
> #Test_1
>
> $ bin/kafka-topics.sh --zookeeper 192.168.1.1:2181 --create --topic
> test1 --partitions 3 --replication-factor 1
> Created topic "test1".
>
> $ time bin/kafka-run-class.sh
> org.apache.kafka.clients.tools.ProducerPerformance test1 10485 100000
> -1 acks=1 bootstrap.servers=192.168.1.1:9092 buffer.memory=67108864
> batch.size=8192
> 10485 records sent, 348.535718 records/sec (33.24 MB/sec)
>
> real    0m30.623s
> user    0m6.704s
> sys     0m1.023s
>
> #Test_2
>
> $bin/kafka-topics.sh --zookeeper 192.168.1.1:2181 --create --topic
> test2 --partitions 1 --replication-factor 1
> Created topic "test2".
>
> $ time bin/kafka-run-class.sh
> org.apache.kafka.clients.tools.ProducerPerformance test1 10485 100000
> -1 acks=1 bootstrap.servers=192.168.1.1:9092 buffer.memory=67108864
> batch.size=8192
> 10485 records sent, 1086.303357 records/sec (103.60 MB/sec)
>
> real    0m10.185s
> user    0m9.138s
> sys     0m1.879s
>
> Literally, #Test_2 ran faster (10 s) and gave better throughput
> (103MB/sec).  The numbers from #Test_1 are just one-thirds that of
> #Test_2.  The topic test1 and test2 differ in such a way that test1
> was created with 3 partitions (each partition sits on different
> physical server) while test2 only has one partition.
>
> My understanding of partition is that more partitions (each on
> different spindle) would lead to better read/write performance because
> of the aggregate bandwidth derived from parallel operations.  But what
> I am seeing here defies my understanding.  I just wonder if I did the
> benchmark tests wrong or I had the concept wrong.
>
> Kind Regards,
>
> Paul
>
> On Tue, Sep 1, 2015 at 12:57 AM, Jay Kreps <jay@confluent.io> wrote:
> > In general increasing message size should increase bytes/sec throughput
> > since much of the work is on a per-message basis. I think the question
> > remains why raising the buffer size with fixed message size would drop
> the
> > throughput. Sounds like a bug if you can reproduce it consistently. Want
> to
> > file a JIRA and see if others can reproduce the same thing?
> >
> > For the multi-server test I may have misread your email. When you say you
> > see 33MB/sec across 3 servers does that mean an aggregate of ~100MB/sec?
> I
> > was assuming yes and what you were seeing was that you were maxing out
> the
> > client's bandwidth so as you added servers each server got a smaller
> chunk
> > of the ~100MB/sec client bandwidth. Maybe that's not what you're saying,
> > though.
> >
> > -Jay
> >
> > On Mon, Aug 31, 2015 at 9:49 AM, explorer <jindotw@gmail.com> wrote:
> >
> >> H Jay,
> >>
> >> Thanks for the response.
> >>
> >> The second command was indeed a typo.  It should have been
> >>
> >> bin/kafka-run-class.sh
> >> org.apache.kafka.clients.tools.ProducerPerformance test1 50000000 100
> >> -1 acks=1 bootstrap.servers=192.168.1.1:9092 buffer.memory=134217728
> >> batch.size=8192
> >>
> >> And the throughput would drop to ~9MB/sec.
> >>
> >> But if I increase the message size, say 10,000 bytes per message
> >>
> >> bin/kafka-run-class.sh
> >> org.apache.kafka.clients.tools.ProducerPerformance test1 500000 10000
> >> -1 acks=1 bootstrap.servers=192.168.1.1:9092 buffer.memory=134217728
> >> batch.size=8192
> >>
> >> The throughput would bounce back to ~33MB/sec.
> >>
> >> I am playing with these numbers just to get a pattern as to what kind of
> >> combination
> >> would serve us the best as far as message size goes.  So it would help
> >> better if we
> >> can safely say that higher buffer memory gives better performance but
> only
> >> to certain extent.
> >>
> >> But in our test context, I get to see lowered throughput with higher
> memory
> >> buffer.  But
> >> once I increase the message size, then the throughput seems normal
> again.
> >> This is the
> >> confusing point.
> >>
> >> For the second part, I am indeed on a 1 gigabit Ethernet.  I just feel
> >> confused why
> >>
> >> A single partition (on a single broker) test yields 100MB/sec throughput
> >>
> >> while
> >>
> >> 3 partitions on 3 brokers (all on different physical server) gave me the
> >> reading of 33MB/sec
> >>
> >> and to make it more clear
> >>
> >> 2 partitions on 2 brokers (on different physical server too) gave me the
> >> reading of 25MB/sec
> >>
> >> I just wanna know how to interpret these numbers so I can draw a pattern
> >> but so far this is
> >> not very consistent (more partitions = less throughput?)
> >>
> >> Cheers,
> >>
> >> Paul
> >>
> >>
> >>
> >> On Tue, Sep 1, 2015 at 12:09 AM, Jay Kreps <jay@confluent.io> wrote:
> >>
> >> > The second command you give actually doesn't seem to double the memory
> >> > (maybe just a typo?). I can't explain why doubling buffer memory would
> >> > decrease throughput. The only effect of adding memory would be if you
> run
> >> > out, and then running out of memory would cause you to block and hence
> >> > lower throughput. So more memory should only be able to help (or have
> no
> >> > effect). I wonder if something else was different between the tests?
> >> >
> >> > For the second test is it possible that you are on 1 gigabit
> ethernet? 1
> >> > gigabit ~= 100mb once you account for the protocol overhead (TCP and
> >> > Kafka's protocol).
> >> >
> >> > -Jay
> >> >
> >> > On Mon, Aug 31, 2015 at 3:14 AM, explorer <jindotw@gmail.com> wrote:
> >> >
> >> > > Hi all,
> >> > >
> >> > > Since my company is considering adopting Kafka as our message bus,
I
> >> > > have been assigned the task to perform some benchmark tests.  I
> >> > > basically followed what Jay wrote on this article
> >> > > <
> >> > >
> >> >
> >>
> http://engineering.linkedin.com/kafka/benchmarking-apache-kafka-2-million-writes-second-three-cheap-machines
> >> > > >
> >> > >
> >> > > The benchmarks were set up using 4 nodes with one node acting as
> both
> >> > > producer and consumer while the rest function as Kafka brokers.
> >> > >
> >> > > This is the baseline (50M messages (100 bytes each) ,64MB buffer
> >> > > memory, and 8192 batch size)
> >> > >
> >> > > bin/kafka-run-class.sh
> >> > > org.apache.kafka.clients.tools.ProducerPerformance test1 50000000
> 100
> >> > > -1 acks=1 bootstrap.servers=192.168.1.1:9092 buffer.memory=67108864
> >> > > batch.size=8192
> >> > >
> >> > > which on our setup yielded the result of
> >> > >
> >> > > 50000000 records sent, 265939.057406 records/sec (25.36 MB/sec)
> >> > >
> >> > > However, by doubling the buffer.memory to 128M
> >> > >
> >> > > bin/kafka-run-class.sh
> >> > > org.apache.kafka.clients.tools.ProducerPerformance test1 500000
> 10000
> >> > > -1 acks=1 bootstrap.servers=192.168.1.1:9092 buffer.memory=67108864
> >> > > batch.size=8192
> >> > >
> >> > > The throughput dropped significantly.
> >> > >
> >> > > 50000000 records sent, 93652.601295 records/sec (8.93 MB/sec)
> >> > >
> >> > > Anyone able to interpret why the throughput degraded so much?
> >> > >
> >> > > Likewise, when performing benchmarks using 3 partitions across 3
> >> > > nodes, the maximum throughput shown is roughly 33.2MB/sec, whereas
a
> >> > > single partition (on a single node) yields 100MB/sec.
> >> > >
> >> > > My guess is that on a 3 nodes setup, I need to multiply the 33.2
> >> > > MB/sec reading by 3 since the the 33.2MB/sec reading only represents
> >> > > the bandwidth available to one single node.
> >> > >
> >> > > Again, anyone out there willing to shed some lights on how to
> >> > > interpret the numbers correctly?
> >> > >
> >> > > Cheers,
> >> > >
> >> > > Paul
> >> > >
> >> >
> >>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message