RE: poor producing performance with very low CPU utilization?
I found one big contributor to the badness was my custom partitioner had a bug
(missing a Utils.ToPositive call).
I also found the default partitioner use of murmur is very bad, compare to
simply doing a hash, for a 5X perf degradation!
As you can see bellow, using a good custom partitioner , I can achieve 1875
message/s, while the default partitioner only gets me to 381 message/s.
Just to share my finding, here are my notes:
public int partition(String topic, Object key, byte[] keyBytes, Object
value, byte[] valueBytes, Cluster cluster) {
if ((keyBytes == null) || (!(key instanceof Long)))
throw new InvalidRecordException("We expect all message
to have a long as key");
return
Utils.toPositive(Utils.murmur2(keyBytes))%cluster.partitionCountForTopic(topic);
}
381 message/s 2.6mb/s
public int partition(String topic, Object key, byte[] keyBytes, Object
value, byte[] valueBytes, Cluster cluster) {
if ((keyBytes == null) || (!(key instanceof Long)))
throw new InvalidRecordException("We expect all message
to have a long as key");
long k = (long) key / 1000; //consecutive 1000 id are co located
return
Utils.murmur2(toBytes(k))%cluster.partitionCountForTopic(topic);
}
private static byte[] toBytes(long val) {
byte [] b = new byte[8];
for (int i = 7; i > 0; i--) {
b[i] = (byte) val;
val >>>= 8;
}
b[0] = (byte) val;
return b;
}
205 message/s 1.4mb/s including some error
public int partition(String topic, Object key, byte[] keyBytes, Object
value, byte[] valueBytes, Cluster cluster) {
if ((keyBytes == null) || (!(key instanceof Long)))
throw new InvalidRecordException("We expect all message
to have a long as key");
long k = (long) key / 1000; //consecutive 1000 id are co located
return
Utils.toPositive(Utils.murmur2(toBytes(k)))%cluster.partitionCountForTopic(topic);
}
381 message/s 2.6mb/s
public int partition(String topic, Object key, byte[] keyBytes,
Object value, byte[] valueBytes, Cluster cluster) {
if ((keyBytes == null) || (!(key instanceof Long)))
throw new InvalidRecordException("We expect all message
to have a long as key");
long k = (long) key / 1000; //consecutive 1000 id are co located
return Long.hashCode(k)%cluster.partitionCountForTopic(topic);
}
1865 message/s 3.7mb/s
public int partition(String topic, Object key, byte[] keyBytes,
Object value, byte[] valueBytes, Cluster cluster) {
if ((keyBytes == null) || (!(key instanceof Long)))
throw new InvalidRecordException("We expect all message
to have a long as key");
long k = (long) key / 1000; //consecutive 1000 id are co located
return (int)(k%cluster.partitionCountForTopic(topic));
1875 message/s 3.7mb/s
-Original Message-
From: Eric Owhadi
Sent: Thursday, October 3, 2019 4:34 PM
To: users@kafka.apache.org
Subject: RE: poor producing performance with very low CPU utilization?
External
To test if my backend was good, I tried using the kafka-producer-perf-test. Boy
that was fast!
Instead of my lame 200 message per seconds, I am getting 20 000 message per
seconds. 100X
That is more in line with my expectations. Granted that this test does not use
my custom partitioner and serializer. I will try add variables one after the
other, but definitelly the bottleneck is not the server :-).
kafka-producer-perf-test --num-records 600 --record-size 12016 --topic
DEFAULT-.TIMESERIES.SmartpumpCollectorVector --throughput 100
--print-metrics --producer-props
bootstrap.servers=nap052.esgyn.local:9092,localhost:9092
compression.type=snappy batch.size=65536 acks=all linger.ms=85
SLF4J: Class path contains multiple SLF4J bindings.
SLF4J: Found binding in
[jar:file:/opt/cloudera/parcels/CDH-5.15.0-1.cdh5.15.0.p0.21/jars/slf4j-log4j12-1.7.5.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: Found binding in
[jar:file:/opt/cloudera/parcels/ESGYNDB-2.7.0-A1/traf_home/export/lib/orc-tools-1.5.0-uber.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: Found binding in
[jar:file:/opt/cloudera/parcels/KAFKA-4.1.0-1.4.1.0.p0.4/lib/kafka/libs/slf4j-log4j12-1.7.5.jar!/org/slf4j/impl/StaticLoggerBinder.class]
SLF4J: See http://www.slf4j.org/codes.html#multiple_bindings for an explanation.
SLF4J: Actual binding is of type [org.slf4j.impl.Log4jLoggerFactory]
19/10/03 14:19:33 INFO producer.ProducerConfig: Produc
RE: poor producing performance with very low CPU utilization?
}
: 61.272
producer-node-metrics:response-total:{client-id=producer-1, node-id=node--1}
: 2.000
producer-node-metrics:response-total:{client-id=producer-1, node-id=node--2}
: 0.000
producer-node-metrics:response-total:{client-id=producer-1, node-id=node-357}
: 18232.000
producer-node-metrics:response-total:{client-id=producer-1, node-id=node-358}
: 18021.000
producer-topic-metrics:byte-rate:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 106864118.709
producer-topic-metrics:byte-total:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 33980657526.000
producer-topic-metrics:compression-rate:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 0.605
producer-topic-metrics:record-error-rate:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 0.000
producer-topic-metrics:record-error-total:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 0.000
producer-topic-metrics:record-retry-rate:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 0.000
producer-topic-metrics:record-retry-total:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 0.000
producer-topic-metrics:record-send-rate:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 18598.670
producer-topic-metrics:record-send-total:{client-id=producer-1,
topic=DEFAULT-.TIMESERIES.SmartpumpCollectorVector} : 600.000
19/10/03 14:24:16 INFO producer.KafkaProducer: [Producer clientId=producer-1]
Closing the Kafka producer with timeoutMillis = 9223372036854775807 ms.
-Original Message-
From: Eric Owhadi
Sent: Thursday, October 3, 2019 3:45 PM
To: users@kafka.apache.org
Subject: RE: poor producing performance with very low CPU utilization?
External
There is a key piece of information that should be critical to guess where the
problem is:
When I change from ack = all to ack = 1, instead of increasing message/s, it
actually devises it by half!
As if the problem is about how fast I produce data (given when I use ack 1 I
assume I block less time in the synchronous send, and therefore my producing
pump increases.
I wonder if some sort of contention happen when producer populate the 200
partition queues when the rate of production is high in the user thread?
Eric
-Original Message-
From: Eric Owhadi
Sent: Thursday, October 3, 2019 1:33 PM
To: users@kafka.apache.org
Subject: RE: poor producing performance with very low CPU utilization?
External
Hi Eric,
Thanks a lot for your answer. Please find inline responses:
>>You've given hardware information about your brokers, but I don't think
>>you've provided information about the machine your producer is running on.
>>Have you verified that you're not reaching any caps on your producer's
>>machine?
The producer is on the same machine that the broker. Running very quiet, 3% CPU
when I run my test. So no there is no stress on the producing side
>>I also think you might be hitting the limit of what a single producer is
>>capable of pushing through with your current setup. With record size of ~12k
>>and the >>default batch size configuration of 64k, you'll only be able to
>>send 5 records per batch. The default number of in flight batches is 5.
I have 200 partition on my topic, and the load is well balanced across all
partition. So the math you are doing should be X200 right? In addition, I found
that batch size had no effect, and the linger.ms was the triggering factor to
cause a buffer send. I played with batch size and in flight number of request
upward, and that had no effect.
>>This means at any given time, you'll only have 25 records in flight per
>>connection. I'm assuming your partitions are configured with at least 2
>>replicas. Acks=all >>means your producer is going to wait for the records to
>>be fully replicated before considering it complete.
>>Doing the math, you have ~200 records per second, but this is split
>>between
>>2 brokers. This means you're producing 100 records per second per broker.
>>Simplifying a bit to 25 records in flight per broker, that's a latency
>>of
>>~250 ms to move around 300kb. At minimum, this includes the time to,
>>[compress the batch], [send the batch over the network to the leader], [write
>>the batch >>to the leader's log], [fetch the batch over the network to the
>>replica], [write the batch to the replica's log], and all of the assorted
>>responses to those calls.
given all is local (producer running on same node as broker), and the size of
my node (
Re: poor producing performance with very low CPU utilization?
Hi Eric, I had the similar issue on my application. One thing I noticed is, metric `kafka_consumer_io_wait_time_avg_seconds` is switching from 2 mins to 6 hours! Average is about 1-2 hours. This metric is a build-in micrometer metric. I'm still googling the source code of this metric to understand better of this metric. Thanks On Thu, Oct 3, 2019 at 4:54 PM Alexandru Ionita wrote: > > This might help. > > Try to replicate the configuration this guy is using for benchmarking kafka. > https://engineering.linkedin.com/kafka/benchmarking-apache-kafka-2-million-writes-second-three-cheap-machines > > Am Do., 3. Okt. 2019 um 22:45 Uhr schrieb Eric Owhadi >: > > > There is a key piece of information that should be critical to guess where > > the problem is: > > > > When I change from ack = all to ack = 1, instead of increasing message/s, > > it actually devises it by half! > > > > As if the problem is about how fast I produce data (given when I use ack 1 > > I assume I block less time in the synchronous send, and therefore my > > producing pump increases. > > > > I wonder if some sort of contention happen when producer populate the 200 > > partition queues when the rate of production is high in the user thread? > > Eric > > > > -Original Message- > > From: Eric Owhadi > > Sent: Thursday, October 3, 2019 1:33 PM > > To: users@kafka.apache.org > > Subject: RE: poor producing performance with very low CPU utilization? > > > > External > > > > Hi Eric, > > Thanks a lot for your answer. Please find inline responses: > > > > >>You've given hardware information about your brokers, but I don't think > > you've provided information about the machine your producer is running on. > > >>Have you verified that you're not reaching any caps on your producer's > > machine? > > > > The producer is on the same machine that the broker. Running very quiet, > > 3% CPU when I run my test. So no there is no stress on the producing side > > > > >>I also think you might be hitting the limit of what a single producer is > > capable of pushing through with your current setup. With record size of > > ~12k and the >>default batch size configuration of 64k, you'll only be able > > to send 5 records per batch. The default number of in flight batches is 5. > > > > I have 200 partition on my topic, and the load is well balanced across all > > partition. So the math you are doing should be X200 right? In addition, I > > found that batch size had no effect, and the linger.ms was the triggering > > factor to cause a buffer send. I played with batch size and in flight > > number of request upward, and that had no effect. > > > > >>This means at any given time, you'll only have 25 records in flight per > > connection. I'm assuming your partitions are configured with at least 2 > > replicas. Acks=all >>means your producer is going to wait for the records > > to be fully replicated before considering it complete. > > > > >>Doing the math, you have ~200 records per second, but this is split > > >>between > > >>2 brokers. This means you're producing 100 records per second per broker. > > >>Simplifying a bit to 25 records in flight per broker, that's a latency > > >>of > > >>~250 ms to move around 300kb. At minimum, this includes the time to, > > [compress the batch], [send the batch over the network to the leader], > > [write the batch >>to the leader's log], [fetch the batch over the network > > to the replica], [write the batch to the replica's log], and all of the > > assorted responses to those calls. > > > > given all is local (producer running on same node as broker), and the size > > of my node (80 vcore), I hope I don t need 250ms to do that... > > The equivalent workload on hbase2.0 is 10 to 20X faster (and that include > > same replica config etc). > > > > On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: > > > > > > -Original Message- > > From: Eric Azama > > Sent: Thursday, October 3, 2019 1:07 PM > > To: users@kafka.apache.org > > Subject: Re: poor producing performance with very low CPU utilization? > > > > External > > > > Hi Eric, > > > > You've given hardware information about your brokers, but I don't think > > you've provided information about the machine your producer is running on. > > Have you verified that you're not reaching any caps on your producer's > > machine? > > > > I also think you might be hitting
Re: poor producing performance with very low CPU utilization?
This might help. Try to replicate the configuration this guy is using for benchmarking kafka. https://engineering.linkedin.com/kafka/benchmarking-apache-kafka-2-million-writes-second-three-cheap-machines Am Do., 3. Okt. 2019 um 22:45 Uhr schrieb Eric Owhadi : > There is a key piece of information that should be critical to guess where > the problem is: > > When I change from ack = all to ack = 1, instead of increasing message/s, > it actually devises it by half! > > As if the problem is about how fast I produce data (given when I use ack 1 > I assume I block less time in the synchronous send, and therefore my > producing pump increases. > > I wonder if some sort of contention happen when producer populate the 200 > partition queues when the rate of production is high in the user thread? > Eric > > -Original Message- > From: Eric Owhadi > Sent: Thursday, October 3, 2019 1:33 PM > To: users@kafka.apache.org > Subject: RE: poor producing performance with very low CPU utilization? > > External > > Hi Eric, > Thanks a lot for your answer. Please find inline responses: > > >>You've given hardware information about your brokers, but I don't think > you've provided information about the machine your producer is running on. > >>Have you verified that you're not reaching any caps on your producer's > machine? > > The producer is on the same machine that the broker. Running very quiet, > 3% CPU when I run my test. So no there is no stress on the producing side > > >>I also think you might be hitting the limit of what a single producer is > capable of pushing through with your current setup. With record size of > ~12k and the >>default batch size configuration of 64k, you'll only be able > to send 5 records per batch. The default number of in flight batches is 5. > > I have 200 partition on my topic, and the load is well balanced across all > partition. So the math you are doing should be X200 right? In addition, I > found that batch size had no effect, and the linger.ms was the triggering > factor to cause a buffer send. I played with batch size and in flight > number of request upward, and that had no effect. > > >>This means at any given time, you'll only have 25 records in flight per > connection. I'm assuming your partitions are configured with at least 2 > replicas. Acks=all >>means your producer is going to wait for the records > to be fully replicated before considering it complete. > > >>Doing the math, you have ~200 records per second, but this is split > >>between > >>2 brokers. This means you're producing 100 records per second per broker. > >>Simplifying a bit to 25 records in flight per broker, that's a latency > >>of > >>~250 ms to move around 300kb. At minimum, this includes the time to, > [compress the batch], [send the batch over the network to the leader], > [write the batch >>to the leader's log], [fetch the batch over the network > to the replica], [write the batch to the replica's log], and all of the > assorted responses to those calls. > > given all is local (producer running on same node as broker), and the size > of my node (80 vcore), I hope I don t need 250ms to do that... > The equivalent workload on hbase2.0 is 10 to 20X faster (and that include > same replica config etc). > > On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: > > > -Original Message- > From: Eric Azama > Sent: Thursday, October 3, 2019 1:07 PM > To: users@kafka.apache.org > Subject: Re: poor producing performance with very low CPU utilization? > > External > > Hi Eric, > > You've given hardware information about your brokers, but I don't think > you've provided information about the machine your producer is running on. > Have you verified that you're not reaching any caps on your producer's > machine? > > I also think you might be hitting the limit of what a single producer is > capable of pushing through with your current setup. With record size of > ~12k and the default batch size configuration of 64k, you'll only be able > to send 5 records per batch. The default number of in flight batches is 5. > This means at any given time, you'll only have 25 records in flight per > connection. I'm assuming your partitions are configured with at least 2 > replicas. Acks=all means your producer is going to wait for the records to > be fully replicated before considering it complete. > > Doing the math, you have ~200 records per second, but this is split between > 2 brokers. This means you're producing 100 records per second per broker. > Simplifying a bit to 25 records in flight per broker, that's a latency of > ~250 ms to move around 300kb. A
RE: poor producing performance with very low CPU utilization?
There is a key piece of information that should be critical to guess where the problem is: When I change from ack = all to ack = 1, instead of increasing message/s, it actually devises it by half! As if the problem is about how fast I produce data (given when I use ack 1 I assume I block less time in the synchronous send, and therefore my producing pump increases. I wonder if some sort of contention happen when producer populate the 200 partition queues when the rate of production is high in the user thread? Eric -Original Message- From: Eric Owhadi Sent: Thursday, October 3, 2019 1:33 PM To: users@kafka.apache.org Subject: RE: poor producing performance with very low CPU utilization? External Hi Eric, Thanks a lot for your answer. Please find inline responses: >>You've given hardware information about your brokers, but I don't think >>you've provided information about the machine your producer is running on. >>Have you verified that you're not reaching any caps on your producer's >>machine? The producer is on the same machine that the broker. Running very quiet, 3% CPU when I run my test. So no there is no stress on the producing side >>I also think you might be hitting the limit of what a single producer is >>capable of pushing through with your current setup. With record size of ~12k >>and the >>default batch size configuration of 64k, you'll only be able to >>send 5 records per batch. The default number of in flight batches is 5. I have 200 partition on my topic, and the load is well balanced across all partition. So the math you are doing should be X200 right? In addition, I found that batch size had no effect, and the linger.ms was the triggering factor to cause a buffer send. I played with batch size and in flight number of request upward, and that had no effect. >>This means at any given time, you'll only have 25 records in flight per >>connection. I'm assuming your partitions are configured with at least 2 >>replicas. Acks=all >>means your producer is going to wait for the records to >>be fully replicated before considering it complete. >>Doing the math, you have ~200 records per second, but this is split >>between >>2 brokers. This means you're producing 100 records per second per broker. >>Simplifying a bit to 25 records in flight per broker, that's a latency >>of >>~250 ms to move around 300kb. At minimum, this includes the time to, >>[compress the batch], [send the batch over the network to the leader], [write >>the batch >>to the leader's log], [fetch the batch over the network to the >>replica], [write the batch to the replica's log], and all of the assorted >>responses to those calls. given all is local (producer running on same node as broker), and the size of my node (80 vcore), I hope I don t need 250ms to do that... The equivalent workload on hbase2.0 is 10 to 20X faster (and that include same replica config etc). On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: -Original Message- From: Eric Azama Sent: Thursday, October 3, 2019 1:07 PM To: users@kafka.apache.org Subject: Re: poor producing performance with very low CPU utilization? External Hi Eric, You've given hardware information about your brokers, but I don't think you've provided information about the machine your producer is running on. Have you verified that you're not reaching any caps on your producer's machine? I also think you might be hitting the limit of what a single producer is capable of pushing through with your current setup. With record size of ~12k and the default batch size configuration of 64k, you'll only be able to send 5 records per batch. The default number of in flight batches is 5. This means at any given time, you'll only have 25 records in flight per connection. I'm assuming your partitions are configured with at least 2 replicas. Acks=all means your producer is going to wait for the records to be fully replicated before considering it complete. Doing the math, you have ~200 records per second, but this is split between 2 brokers. This means you're producing 100 records per second per broker. Simplifying a bit to 25 records in flight per broker, that's a latency of ~250 ms to move around 300kb. At minimum, this includes the time to, [compress the batch], [send the batch over the network to the leader], [write the batch to the leader's log], [fetch the batch over the network to the replica], [write the batch to the replica's log], and all of the assorted responses to those calls. On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: > Hi Jamie, > Thanks for the hint. I played with these parameters, and found only > linger.ms is playing a significant role for my test case. > It is very sensitive and highly non linear. > I get these results: > Linger.ms
RE: poor producing performance with very low CPU utilization?
Hi Eric, Thanks a lot for your answer. Please find inline responses: >>You've given hardware information about your brokers, but I don't think >>you've provided information about the machine your producer is running on. >>Have you verified that you're not reaching any caps on your producer's >>machine? The producer is on the same machine that the broker. Running very quiet, 3% CPU when I run my test. So no there is no stress on the producing side >>I also think you might be hitting the limit of what a single producer is >>capable of pushing through with your current setup. With record size of ~12k >>and the >>default batch size configuration of 64k, you'll only be able to >>send 5 records per batch. The default number of in flight batches is 5. I have 200 partition on my topic, and the load is well balanced across all partition. So the math you are doing should be X200 right? In addition, I found that batch size had no effect, and the linger.ms was the triggering factor to cause a buffer send. I played with batch size and in flight number of request upward, and that had no effect. >>This means at any given time, you'll only have 25 records in flight per >>connection. I'm assuming your partitions are configured with at least 2 >>replicas. Acks=all >>means your producer is going to wait for the records to >>be fully replicated before considering it complete. >>Doing the math, you have ~200 records per second, but this is split between >>2 brokers. This means you're producing 100 records per second per broker. >>Simplifying a bit to 25 records in flight per broker, that's a latency of >>~250 ms to move around 300kb. At minimum, this includes the time to, >>[compress the batch], [send the batch over the network to the leader], [write >>the batch >>to the leader's log], [fetch the batch over the network to the >>replica], [write the batch to the replica's log], and all of the assorted >>responses to those calls. given all is local (producer running on same node as broker), and the size of my node (80 vcore), I hope I don t need 250ms to do that... The equivalent workload on hbase2.0 is 10 to 20X faster (and that include same replica config etc). On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: -----Original Message----- From: Eric Azama Sent: Thursday, October 3, 2019 1:07 PM To: users@kafka.apache.org Subject: Re: poor producing performance with very low CPU utilization? External Hi Eric, You've given hardware information about your brokers, but I don't think you've provided information about the machine your producer is running on. Have you verified that you're not reaching any caps on your producer's machine? I also think you might be hitting the limit of what a single producer is capable of pushing through with your current setup. With record size of ~12k and the default batch size configuration of 64k, you'll only be able to send 5 records per batch. The default number of in flight batches is 5. This means at any given time, you'll only have 25 records in flight per connection. I'm assuming your partitions are configured with at least 2 replicas. Acks=all means your producer is going to wait for the records to be fully replicated before considering it complete. Doing the math, you have ~200 records per second, but this is split between 2 brokers. This means you're producing 100 records per second per broker. Simplifying a bit to 25 records in flight per broker, that's a latency of ~250 ms to move around 300kb. At minimum, this includes the time to, [compress the batch], [send the batch over the network to the leader], [write the batch to the leader's log], [fetch the batch over the network to the replica], [write the batch to the replica's log], and all of the assorted responses to those calls. On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: > Hi Jamie, > Thanks for the hint. I played with these parameters, and found only > linger.ms is playing a significant role for my test case. > It is very sensitive and highly non linear. > I get these results: > Linger.ms message per second > 80 100 > 84 205 > 85 215 -> top > 86 213 > 90 205 > 95 195 > 100 187 > 200 100 > > So as you can see, this is very sensitive and one can miss the peek easily. > However, 200 messages per second for 2 powerful nodes and relatively > small message (12016bytes) is still at least 10X bellow what I would have > hoped. > When I see system resources still being barely moving, with cpu at 3%, > I am sure something is not right. > Regards, > Eric > > -Original Message- > From: Jamie > Sent: Wednesday, October 2, 2019 4:27 PM > To: users
Re: poor producing performance with very low CPU utilization?
Hi Eric, You've given hardware information about your brokers, but I don't think you've provided information about the machine your producer is running on. Have you verified that you're not reaching any caps on your producer's machine? I also think you might be hitting the limit of what a single producer is capable of pushing through with your current setup. With record size of ~12k and the default batch size configuration of 64k, you'll only be able to send 5 records per batch. The default number of in flight batches is 5. This means at any given time, you'll only have 25 records in flight per connection. I'm assuming your partitions are configured with at least 2 replicas. Acks=all means your producer is going to wait for the records to be fully replicated before considering it complete. Doing the math, you have ~200 records per second, but this is split between 2 brokers. This means you're producing 100 records per second per broker. Simplifying a bit to 25 records in flight per broker, that's a latency of ~250 ms to move around 300kb. At minimum, this includes the time to, [compress the batch], [send the batch over the network to the leader], [write the batch to the leader's log], [fetch the batch over the network to the replica], [write the batch to the replica's log], and all of the assorted responses to those calls. On Wed, Oct 2, 2019 at 8:38 PM Eric Owhadi wrote: > Hi Jamie, > Thanks for the hint. I played with these parameters, and found only > linger.ms is playing a significant role for my test case. > It is very sensitive and highly non linear. > I get these results: > Linger.ms message per second > 80 100 > 84 205 > 85 215 -> top > 86 213 > 90 205 > 95 195 > 100 187 > 200 100 > > So as you can see, this is very sensitive and one can miss the peek easily. > However, 200 messages per second for 2 powerful nodes and relatively small > message (12016bytes) is still at least 10X bellow what I would have hoped. > When I see system resources still being barely moving, with cpu at 3%, I am > sure something is not right. > Regards, > Eric > > -Original Message- > From: Jamie > Sent: Wednesday, October 2, 2019 4:27 PM > To: users@kafka.apache.org > Subject: Re: poor producing performance with very low CPU utilization? > > External > > Hi Eric, > I found increasing the linger.ms to between 50-100 ms significantly > increases performance (fewer larger requests instead of many small ones), > I'd also increase the batch size and the buffer.memory. > Thanks, > Jamie > > > -Original Message- > From: Eric Owhadi > To: users@kafka.apache.org > Sent: Wed, 2 Oct 2019 16:42 > Subject: poor producing performance with very low CPU utilization? > > Hi Kafka users, > I am new to Kafka and am struggling with getting acceptable producing rate. > I am using a cluster of 2 nodes, 40 CPU cores/ 80 if counting > hyperthreading. 256GB memory on a 10Gbit network Kafka is installed as part > of cloudera parcel, with 5GB java heap. > Producer version: Kafka client 2.2.1 > > Wed Oct 2 07:56:59 PDT 2019 > JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.131-0.b11.el6_9.x86_64 > Using -XX:+HeapDumpOnOutOfMemoryError > -XX:HeapDumpPath=/tmp/kafka_kafka-KAFKA_BROKER-c1871edf37153578a6fc7f41462d01d2_pid6908.hprof > -XX:OnOutOfMemoryError=/usr/lib64/cmf/service/common/killparent.sh as > CSD_JAVA_OPTS Using > /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER as conf dir > Using scripts/control.sh as process script > CONF_DIR=/var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER > CMF_CONF_DIR=/etc/cloudera-scm-agent > > Date: Wed Oct 2 07:56:59 PDT 2019 > Host: x.esgyn.local > Pwd: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER > CONF_DIR: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER > KAFKA_HOME: /opt/cloudera/parcels/KAFKA-4.1.0-1.4.1.0.p0.4/lib/kafka > Zookeeper Quorum: > xxx.esgyn.local:2181,xxx.esgyn.local:2181,xxx.esgyn.local:2181 > Zookeeper Chroot: > PORT: 9092 > JMX_PORT: 9393 > SSL_PORT: 9093 > ENABLE_MONITORING: true > METRIC_REPORTERS: nl.techop.kafka.KafkaHttpMetricsReporter > BROKER_HEAP_SIZE: 5120 > BROKER_JAVA_OPTS: -server -XX:+UseG1GC -XX:MaxGCPauseMillis=20 > -XX:InitiatingHeapOccupancyPercent=35 -XX:+DisableExplicitGC > -Djava.awt.headless=true > BROKER_SSL_ENABLED: false > KERBEROS_AUTH_ENABLED: false > KAFKA_PRINCIPAL: > SECURITY_INTER_BROKER_PROTOCOL: INFERRED > AUTHENTICATE_ZOOKEEPER_CONNECTION: true > SUPER_USERS: kafka > Kafka version found: 2.2.1-kafka4.1.0 > Sentry version found: 1.5.1-cdh5.15.0 > ZK_PRINCIPAL_NAME: zookeeper > Final Zookeeper Quorum is
RE: poor producing performance with very low CPU utilization?
Hi Jamie,
Thanks for the hint. I played with these parameters, and found only linger.ms
is playing a significant role for my test case.
It is very sensitive and highly non linear.
I get these results:
Linger.ms message per second
80 100
84 205
85 215 -> top
86 213
90 205
95 195
100 187
200 100
So as you can see, this is very sensitive and one can miss the peek easily.
However, 200 messages per second for 2 powerful nodes and relatively small
message (12016bytes) is still at least 10X bellow what I would have hoped. When
I see system resources still being barely moving, with cpu at 3%, I am sure
something is not right.
Regards,
Eric
-Original Message-
From: Jamie
Sent: Wednesday, October 2, 2019 4:27 PM
To: users@kafka.apache.org
Subject: Re: poor producing performance with very low CPU utilization?
External
Hi Eric,
I found increasing the linger.ms to between 50-100 ms significantly increases
performance (fewer larger requests instead of many small ones), I'd also
increase the batch size and the buffer.memory.
Thanks,
Jamie
-Original Message-
From: Eric Owhadi
To: users@kafka.apache.org
Sent: Wed, 2 Oct 2019 16:42
Subject: poor producing performance with very low CPU utilization?
Hi Kafka users,
I am new to Kafka and am struggling with getting acceptable producing rate.
I am using a cluster of 2 nodes, 40 CPU cores/ 80 if counting hyperthreading.
256GB memory on a 10Gbit network Kafka is installed as part of cloudera parcel,
with 5GB java heap.
Producer version: Kafka client 2.2.1
Wed Oct 2 07:56:59 PDT 2019
JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.131-0.b11.el6_9.x86_64
Using -XX:+HeapDumpOnOutOfMemoryError
-XX:HeapDumpPath=/tmp/kafka_kafka-KAFKA_BROKER-c1871edf37153578a6fc7f41462d01d2_pid6908.hprof
-XX:OnOutOfMemoryError=/usr/lib64/cmf/service/common/killparent.sh as
CSD_JAVA_OPTS Using
/var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER as conf dir Using
scripts/control.sh as process script
CONF_DIR=/var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
CMF_CONF_DIR=/etc/cloudera-scm-agent
Date: Wed Oct 2 07:56:59 PDT 2019
Host: x.esgyn.local
Pwd: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
CONF_DIR: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
KAFKA_HOME: /opt/cloudera/parcels/KAFKA-4.1.0-1.4.1.0.p0.4/lib/kafka
Zookeeper Quorum: xxx.esgyn.local:2181,xxx.esgyn.local:2181,xxx.esgyn.local:2181
Zookeeper Chroot:
PORT: 9092
JMX_PORT: 9393
SSL_PORT: 9093
ENABLE_MONITORING: true
METRIC_REPORTERS: nl.techop.kafka.KafkaHttpMetricsReporter
BROKER_HEAP_SIZE: 5120
BROKER_JAVA_OPTS: -server -XX:+UseG1GC -XX:MaxGCPauseMillis=20
-XX:InitiatingHeapOccupancyPercent=35 -XX:+DisableExplicitGC
-Djava.awt.headless=true
BROKER_SSL_ENABLED: false
KERBEROS_AUTH_ENABLED: false
KAFKA_PRINCIPAL:
SECURITY_INTER_BROKER_PROTOCOL: INFERRED
AUTHENTICATE_ZOOKEEPER_CONNECTION: true
SUPER_USERS: kafka
Kafka version found: 2.2.1-kafka4.1.0
Sentry version found: 1.5.1-cdh5.15.0
ZK_PRINCIPAL_NAME: zookeeper
Final Zookeeper Quorum is
xxx.esgyn.local:2181,xx.esgyn.local:2181,x.esgyn.local:2181
security.inter.broker.protocol inferred as PLAINTEXT
LISTENERS=listeners=PLAINTEXT://x.esgyn.local:9092,
I am producing messages of 12016 bytes uncompressed, then snappy compressed by
kafka.
I am using a topic with 200 partitions, and a custom partitioner that I
verified is doing good job at spreading the load on the 2 brokers.
My producer config look like:
kafkaProps.put("bootstrap.servers","nap052.esgyn.local:9092,localhost:9092");
kafkaProps.put("key.serializer",
"org.apache.kafka.common.serialization.LongSerializer");
kafkaProps.put("value.serializer","org.trafodion.sql.kafka.SmartpumpCollectorVectorSerializer");
kafkaProps.put("partitioner.class","org.trafodion.sql.kafka.TimeSeriesPartitioner");
kafkaProps.put("compression.type","snappy");
kafkaProps.put("batch.size","65536");
kafkaProps.put("acks", "all");
kafkaProps.put("linger.ms","1");
I tried first doing fire and forget send, thinking I would get best performance.
Then I tried synchronous send, and amazingly found that I would get better
performance with sync send.
However, after 1 or 2 minute of load test, I start getting error on the
synchronous send like this:
ava.util.concurrent.ExecutionException:
org.apache.kafka.common.errors.TimeoutException: Expiring 1 record(s) for
DEFAULT-.TIMESERIES.SmartpumpCollectorVector--112:12 ms has passed since
batch creation
at
org.apache.kafka.clients.producer.internals.FutureReco
Re: poor producing performance with very low CPU utilization?
Hi Eric,
I found increasing the linger.ms to between 50-100 ms significantly increases
performance (fewer larger requests instead of many small ones), I'd also
increase the batch size and the buffer.memory.
Thanks,
Jamie
-Original Message-
From: Eric Owhadi
To: users@kafka.apache.org
Sent: Wed, 2 Oct 2019 16:42
Subject: poor producing performance with very low CPU utilization?
Hi Kafka users,
I am new to Kafka and am struggling with getting acceptable producing rate.
I am using a cluster of 2 nodes, 40 CPU cores/ 80 if counting hyperthreading.
256GB memory on a 10Gbit network
Kafka is installed as part of cloudera parcel, with 5GB java heap.
Producer version: Kafka client 2.2.1
Wed Oct 2 07:56:59 PDT 2019
JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.131-0.b11.el6_9.x86_64
Using -XX:+HeapDumpOnOutOfMemoryError
-XX:HeapDumpPath=/tmp/kafka_kafka-KAFKA_BROKER-c1871edf37153578a6fc7f41462d01d2_pid6908.hprof
-XX:OnOutOfMemoryError=/usr/lib64/cmf/service/common/killparent.sh as
CSD_JAVA_OPTS
Using /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER as conf dir
Using scripts/control.sh as process script
CONF_DIR=/var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
CMF_CONF_DIR=/etc/cloudera-scm-agent
Date: Wed Oct 2 07:56:59 PDT 2019
Host: x.esgyn.local
Pwd: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
CONF_DIR: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
KAFKA_HOME: /opt/cloudera/parcels/KAFKA-4.1.0-1.4.1.0.p0.4/lib/kafka
Zookeeper Quorum: xxx.esgyn.local:2181,xxx.esgyn.local:2181,xxx.esgyn.local:2181
Zookeeper Chroot:
PORT: 9092
JMX_PORT: 9393
SSL_PORT: 9093
ENABLE_MONITORING: true
METRIC_REPORTERS: nl.techop.kafka.KafkaHttpMetricsReporter
BROKER_HEAP_SIZE: 5120
BROKER_JAVA_OPTS: -server -XX:+UseG1GC -XX:MaxGCPauseMillis=20
-XX:InitiatingHeapOccupancyPercent=35 -XX:+DisableExplicitGC
-Djava.awt.headless=true
BROKER_SSL_ENABLED: false
KERBEROS_AUTH_ENABLED: false
KAFKA_PRINCIPAL:
SECURITY_INTER_BROKER_PROTOCOL: INFERRED
AUTHENTICATE_ZOOKEEPER_CONNECTION: true
SUPER_USERS: kafka
Kafka version found: 2.2.1-kafka4.1.0
Sentry version found: 1.5.1-cdh5.15.0
ZK_PRINCIPAL_NAME: zookeeper
Final Zookeeper Quorum is
xxx.esgyn.local:2181,xx.esgyn.local:2181,x.esgyn.local:2181
security.inter.broker.protocol inferred as PLAINTEXT
LISTENERS=listeners=PLAINTEXT://x.esgyn.local:9092,
I am producing messages of 12016 bytes uncompressed, then snappy compressed by
kafka.
I am using a topic with 200 partitions, and a custom partitioner that I
verified is doing good job at spreading the load on the 2 brokers.
My producer config look like:
kafkaProps.put("bootstrap.servers","nap052.esgyn.local:9092,localhost:9092");
kafkaProps.put("key.serializer",
"org.apache.kafka.common.serialization.LongSerializer");
kafkaProps.put("value.serializer","org.trafodion.sql.kafka.SmartpumpCollectorVectorSerializer");
kafkaProps.put("partitioner.class","org.trafodion.sql.kafka.TimeSeriesPartitioner");
kafkaProps.put("compression.type","snappy");
kafkaProps.put("batch.size","65536");
kafkaProps.put("acks", "all");
kafkaProps.put("linger.ms","1");
I tried first doing fire and forget send, thinking I would get best performance.
Then I tried synchronous send, and amazingly found that I would get better
performance with sync send.
However, after 1 or 2 minute of load test, I start getting error on the
synchronous send like this:
ava.util.concurrent.ExecutionException:
org.apache.kafka.common.errors.TimeoutException: Expiring 1 record(s) for
DEFAULT-.TIMESERIES.SmartpumpCollectorVector--112:12 ms has passed since
batch creation
at
org.apache.kafka.clients.producer.internals.FutureRecordMetadata.valueOrError(FutureRecordMetadata.java:98)
at
org.apache.kafka.clients.producer.internals.FutureRecordMetadata.get(FutureRecordMetadata.java:67)
at
org.apache.kafka.clients.producer.internals.FutureRecordMetadata.get(FutureRecordMetadata.java:30)
at
org.trafodion.sql.kafka.TimeseriesEndPoint$customHandler.handle(TimeseriesEndPoint.java:315)
at
org.eclipse.jetty.server.handler.HandlerWrapper.handle(HandlerWrapper.java:132)
at org.eclipse.jetty.server.Server.handle(Server.java:505)
at org.eclipse.jetty.server.HttpChannel.handle(HttpChannel.java:370)
at
org.eclipse.jetty.server.HttpConnection.onFillable(HttpConnection.java:267)
at
org.eclipse.jetty.io.AbstractConnection$ReadCallback.succeeded(AbstractConnection.java:305)
at org.eclipse.jetty.io.FillInterest.fillable(FillInterest.java:103)
at org.ec
poor producing performance with very low CPU utilization?
Hi Kafka users,
I am new to Kafka and am struggling with getting acceptable producing rate.
I am using a cluster of 2 nodes, 40 CPU cores/ 80 if counting hyperthreading.
256GB memory on a 10Gbit network
Kafka is installed as part of cloudera parcel, with 5GB java heap.
Producer version: Kafka client 2.2.1
Wed Oct 2 07:56:59 PDT 2019
JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.131-0.b11.el6_9.x86_64
Using -XX:+HeapDumpOnOutOfMemoryError
-XX:HeapDumpPath=/tmp/kafka_kafka-KAFKA_BROKER-c1871edf37153578a6fc7f41462d01d2_pid6908.hprof
-XX:OnOutOfMemoryError=/usr/lib64/cmf/service/common/killparent.sh as
CSD_JAVA_OPTS
Using /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER as conf dir
Using scripts/control.sh as process script
CONF_DIR=/var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
CMF_CONF_DIR=/etc/cloudera-scm-agent
Date: Wed Oct 2 07:56:59 PDT 2019
Host: x.esgyn.local
Pwd: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
CONF_DIR: /var/run/cloudera-scm-agent/process/33853-kafka-KAFKA_BROKER
KAFKA_HOME: /opt/cloudera/parcels/KAFKA-4.1.0-1.4.1.0.p0.4/lib/kafka
Zookeeper Quorum: xxx.esgyn.local:2181,xxx.esgyn.local:2181,xxx.esgyn.local:2181
Zookeeper Chroot:
PORT: 9092
JMX_PORT: 9393
SSL_PORT: 9093
ENABLE_MONITORING: true
METRIC_REPORTERS: nl.techop.kafka.KafkaHttpMetricsReporter
BROKER_HEAP_SIZE: 5120
BROKER_JAVA_OPTS: -server -XX:+UseG1GC -XX:MaxGCPauseMillis=20
-XX:InitiatingHeapOccupancyPercent=35 -XX:+DisableExplicitGC
-Djava.awt.headless=true
BROKER_SSL_ENABLED: false
KERBEROS_AUTH_ENABLED: false
KAFKA_PRINCIPAL:
SECURITY_INTER_BROKER_PROTOCOL: INFERRED
AUTHENTICATE_ZOOKEEPER_CONNECTION: true
SUPER_USERS: kafka
Kafka version found: 2.2.1-kafka4.1.0
Sentry version found: 1.5.1-cdh5.15.0
ZK_PRINCIPAL_NAME: zookeeper
Final Zookeeper Quorum is
xxx.esgyn.local:2181,xx.esgyn.local:2181,x.esgyn.local:2181
security.inter.broker.protocol inferred as PLAINTEXT
LISTENERS=listeners=PLAINTEXT://x.esgyn.local:9092,
I am producing messages of 12016 bytes uncompressed, then snappy compressed by
kafka.
I am using a topic with 200 partitions, and a custom partitioner that I
verified is doing good job at spreading the load on the 2 brokers.
My producer config look like:
kafkaProps.put("bootstrap.servers","nap052.esgyn.local:9092,localhost:9092");
kafkaProps.put("key.serializer",
"org.apache.kafka.common.serialization.LongSerializer");
kafkaProps.put("value.serializer","org.trafodion.sql.kafka.SmartpumpCollectorVectorSerializer");
kafkaProps.put("partitioner.class","org.trafodion.sql.kafka.TimeSeriesPartitioner");
kafkaProps.put("compression.type","snappy");
kafkaProps.put("batch.size","65536");
kafkaProps.put("acks", "all");
kafkaProps.put("linger.ms","1");
I tried first doing fire and forget send, thinking I would get best performance.
Then I tried synchronous send, and amazingly found that I would get better
performance with sync send.
However, after 1 or 2 minute of load test, I start getting error on the
synchronous send like this:
ava.util.concurrent.ExecutionException:
org.apache.kafka.common.errors.TimeoutException: Expiring 1 record(s) for
DEFAULT-.TIMESERIES.SmartpumpCollectorVector--112:12 ms has passed since
batch creation
at
org.apache.kafka.clients.producer.internals.FutureRecordMetadata.valueOrError(FutureRecordMetadata.java:98)
at
org.apache.kafka.clients.producer.internals.FutureRecordMetadata.get(FutureRecordMetadata.java:67)
at
org.apache.kafka.clients.producer.internals.FutureRecordMetadata.get(FutureRecordMetadata.java:30)
at
org.trafodion.sql.kafka.TimeseriesEndPoint$customHandler.handle(TimeseriesEndPoint.java:315)
at
org.eclipse.jetty.server.handler.HandlerWrapper.handle(HandlerWrapper.java:132)
at org.eclipse.jetty.server.Server.handle(Server.java:505)
at org.eclipse.jetty.server.HttpChannel.handle(HttpChannel.java:370)
at
org.eclipse.jetty.server.HttpConnection.onFillable(HttpConnection.java:267)
at
org.eclipse.jetty.io.AbstractConnection$ReadCallback.succeeded(AbstractConnection.java:305)
at org.eclipse.jetty.io.FillInterest.fillable(FillInterest.java:103)
at org.eclipse.jetty.io.ChannelEndPoint$2.run(ChannelEndPoint.java:117)
at
org.eclipse.jetty.util.thread.strategy.EatWhatYouKill.runTask(EatWhatYouKill.java:333)
at
org.eclipse.jetty.util.thread.strategy.EatWhatYouKill.doProduce(EatWhatYouKill.java:310)
at
org.eclipse.jetty.util.thread.strategy.EatWhatYouKill.tryProduce(EatWhatYouKill.java:168)
at
org.eclipse.jetty.util.thread.strategy.EatWhatYouKill.run(EatWhatYouKill.java:126)
at
org.eclipse.jetty.util.thread.ReservedThreadExecutor$ReservedThread.run(ReservedThreadExecutor.java:366)
at
