kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Prabhjot Bharaj <prabhbha...@gmail.com>
Subject Re: Slow ISR catch-up
Date Fri, 04 Sep 2015 14:39:56 GMT
Hi,

I am experiencing super slow throughput when using acks=-1
Some further progress in continuation to the test in my previous email:-

*Topic details -*

Topic:temp PartitionCount:1 ReplicationFactor:3 Configs:

Topic: temp Partition: 0 Leader: 5 Replicas: 5,1,2 Isr: 5,2,1
*This is the command I'm running - *

kafka-producer-perf-test.sh --broker-list 96.7.250.122:9092,
96.17.183.53:9092,96.17.183.54:9092,96.7.250.117:9092,96.7.250.118:9092
--messages 100000 --message-size 500 --topics temp --show-detailed-stats
--threads 30 --request-num-acks -1 --batch-size 1000 --request-timeout-ms
10000

*Server.properties:-*

broker.id=0

port=9092

*num.network.threads=6*

num.io.threads=8

*socket.send.buffer.bytes=10485760*

*socket.receive.buffer.bytes=10485760*

socket.request.max.bytes=104857600

log.dirs=/tmp/kafka-logs

num.partitions=1

num.recovery.threads.per.data.dir=1

log.retention.hours=168

log.segment.bytes=1073741824

log.retention.check.interval.ms=300000

log.cleaner.enable=false

zookeeper.connect=localhost:2181

zookeeper.connection.timeout.ms=6000

*num.replica.fetchers=6*
*Observation:-*

I have also noticed that if I use acks=1 (without --sync) and immediately
use acks=-1 (without --sync), the test completes very quickly. Also, after
running  this, if I describe the topic, it is still not in sync, which
means acks=-1 is treated as 1 only

Also, when running on a freshly created topic with just acks=-1, it takes 8
minutes to complete with --sync

time kafka-producer-perf-test.sh --broker-list 96.7.250.122:9092,
96.17.183.53:9092,96.17.183.54:9092,96.7.250.117:9092,96.7.250.118:9092
--messages 100000 --message-size 500 --topics temp --show-detailed-stats
--threads 30 --request-num-acks -1 --batch-size 1000 --request-timeout-ms
10000 --compression-codec 2 --sync

start.time, end.time, compression, message.size, batch.size,
total.data.sent.in.MB, MB.sec, total.data.sent.in.nMsg, nMsg.sec

2015-09-04 12:19:36:223, 2015-09-04 12:27:41:775, 2, 500, 1000, 47.68,
0.0982, 99990, 205.9306

real 8m6.563s

user 0m19.787s

sys 0m5.601s

If I use --sync, it is taking way longer.

Where am I doing wrong?

Thanks,
Prabhjot

On Fri, Sep 4, 2015 at 1:45 AM, Gwen Shapira <gwen@confluent.io> wrote:

> Yes, this should work. Expect lower throughput though.
>
> On Thu, Sep 3, 2015 at 12:52 PM, Prabhjot Bharaj <prabhbharaj@gmail.com>
> wrote:
>
> > Hi,
> >
> > Can I use sync for acks = -1?
> >
> > Regards,
> > Prabhjot
> > On Sep 3, 2015 11:49 PM, "Gwen Shapira" <gwen@confluent.io> wrote:
> >
> > > The test uses the old producer (we should fix that), and since you
> don't
> > > specify --sync, it runs async.
> > > The old async producer simply sends data and doesn't wait for acks, so
> it
> > > is possible that the messages were never acked...
> > >
> > > On Thu, Sep 3, 2015 at 7:56 AM, Prabhjot Bharaj <prabhbharaj@gmail.com
> >
> > > wrote:
> > >
> > > > Hi Folks,
> > > >
> > > > Request your expertise on my doubt here.
> > > >
> > > > *My setup:-*
> > > >
> > > > 5 node kafka cluster (4 cores, 8GB RAM) on RAID-6 (500 GB)
> > > > Using Kafka 0.8.2.1 with modified ProducerPerformance.scala
> > > > I've modified ProducerPerformance.scala to send custom ASCII data,
> > > instead
> > > > of Byte Array of Zeroes
> > > >
> > > > *server.properties:-*
> > > >
> > > > broker.id=0
> > > >
> > > > log.cleaner.enable=false
> > > >
> > > > log.dirs=/tmp/kafka-logs
> > > >
> > > > log.retention.check.interval.ms=300000
> > > >
> > > > log.retention.hours=168
> > > >
> > > > log.segment.bytes=1073741824
> > > >
> > > > num.io.threads=8
> > > >
> > > > num.network.threads=3
> > > >
> > > > num.partitions=1
> > > >
> > > > num.recovery.threads.per.data.dir=1
> > > >
> > > > *num.replica.fetchers=4*
> > > >
> > > > port=9092
> > > >
> > > > socket.receive.buffer.bytes=1048576
> > > >
> > > > socket.request.max.bytes=104857600
> > > >
> > > > socket.send.buffer.bytes=1048576
> > > >
> > > > zookeeper.connect=localhost:2181
> > > >
> > > > zookeeper.connection.timeout.ms=6000
> > > >
> > > >
> > > > *This is how I run the producer perf test:-*
> > > >
> > > > kafka-producer-perf-test.sh --broker-list
> > > > a.a.a.a:9092,b.b.b.b:9092,c.c.c.c:9092,d.d.d.d:9092,e.e.e.e:9092
> > > --messages
> > > > 100000 --message-size 500 --topics temp --show-detailed-stats
> > --threads
> > > 5
> > > > --request-num-acks -1 --batch-size 200 --request-timeout-ms 10000
> > > > --compression-codec 0
> > > >
> > > > *Problem:-*
> > > >
> > > > This test completes in under 15 seconds for me
> > > >
> > > > But, after this test, if I try writing to another topic which has 2
> > > > partitions and 3 replicas, it is dead slow and the same script seems
> > > never
> > > > to finish because the slow ISR catch-up is still going on.
> > > >
> > > > *My inference:-*
> > > > I have noticed that for a topic with 1 partition and 3 replicas, the
> > ISR
> > > > shows only 1 broker id.
> > > >
> > > > Topic:temp PartitionCount:1 ReplicationFactor:3 Configs:
> > > >
> > > > Topic: temp Partition: 0 Leader: 5 Replicas: 5,1,2 Isr: 5
> > > >
> > > >
> > > > I think it is because the data from the leader is not received in
> > broker
> > > > ids 1 and 2
> > > > Also, I could confirm it from the data directory sizes for this
> topic.
> > > > Leader (5) has 20GB but replicas - 1 and 2 are still at 7GB
> > > >
> > > > *Doubts:-*
> > > > 1. But, I was running the kafka-producer-perf-test.sh with acks=-1,
> > which
> > > > means that all data must have been committed to all replicas. But,
> with
> > > the
> > > > replicas still at 7GB, it doesnt seem that acks=-1 is considered by
> the
> > > > producer.
> > > >
> > > > Am I missing something ?
> > > >
> > > > Regards,
> > > > Prabhjot
> > > >
> > >
> >
>



-- 
---------------------------------------------------------
"There are only 10 types of people in the world: Those who understand
binary, and those who don't"

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