On 21/04/2016 16:46, Aljoscha Krettek wrote:
Hi,
I would be very happy about improvements to our RocksDB performance.
What are the RocksDB Java benchmarks that you are running? In Flink,
we also have to serialize/deserialize every time that we access
RocksDB using our TypeSerializer. Maybe this is causing the slow down.
Hi Aljoscha,
I'm using benchmark from:
https://github.com/facebook/rocksdb/blob/master/java/jdb_bench.sh
My value is pretty simple scala case class - around 12 fields with
Int/Long/String values - I think serialization shouldn't be a big
problem. However I think I'll have to do more comprehensive tests to be
sure I'm comparing apples to apples - hope to find time during weekend
for that :)
thanks,
maciek
By the way, what is the type of value stored in the RocksDB state.
Maybe the TypeSerializer for that value is very slow.
Cheers,
Aljoscha
On Thu, 21 Apr 2016 at 16:41 Maciek Próchniak <m...@touk.pl
<mailto:m...@touk.pl>> wrote:
Well...
I found some time to look at rocksDB performance.
It takes around 0.4ms to lookup value state and 0.12ms to update -
these are means, 95th percentile was > 1ms for get... When I set
additional options:
.setIncreaseParallelism(8)
.setMaxOpenFiles(-1)
.setCompressionType(CompressionType.SNAPPY_COMPRESSION)
I manage to get
0.05ms for update and 0.2ms for get - but still it seems pretty
bad compared to standard rocksdb java benchmarks that I try on the
same machine, as they are:
fillseq : 1.23238 micros/op; 89.8 MB/s; 1000000 ops
done; 1 / 1 task(s) finished.
readrandom : 9.25380 micros/op; 12.0 MB/s; 1000000 /
1000000 found; 1 / 1 task(s) finished.
fillrandom : 4.46839 micros/op; 24.8 MB/s; 1000000 ops
done; 1 / 1 task(s) finished.
guess I'll have to look at it a bit more...
thanks anyway,
maciek
On 21/04/2016 08:41, Maciek Próchniak wrote:
Hi Ufuk,
thanks for quick reply.
Actually I had a little time to try both things.
1) helped only temporarily - it just took a bit longer to
saturate the pool. After few minutes, periodically all kafka
threads were waiting for bufferPool.
2) This seemed to help. I also reduced checkpoint interval - on
rocks we had 5min, now I tried 30s. .
I attach throughput metrics - the former (around 18) is with
increased heap & buffers, the latter (around 22) is with
FileSystemStateBackend.
My state is few GB large - during the test it reached around
2-3GB. I must admit I was quite impressed that checkpointing to
HDFS using FileSystem took only about 6-7s (with occasional
spikes to 12-13s, which can be seen on metrcs - didn't check if
it was caused by hdfs or sth else).
Now I looked at logs from 18 and seems like checkpointing rocksdb
took around 2-3minutes:
2016-04-20 17:47:33,439 [Checkpoint Timer] INFO
org.apache.flink.runtime.checkpoint.CheckpointCoordinator -
Triggering checkpoint 6 @ 1461167253439
2016-04-20 17:49:54,196 [flink-akka.actor.default-dispatcher-147]
INFO org.apache.flink.runtime.checkpoint.CheckpointCoordinator -
Completed checkpoint 6 (in 140588 ms)
- however I don't see any threads dumping state in threadStacks...
I guess I'll have to add some metrics around state invocations to
see where is the problem with rocksDB... I'll write if I find
anything, but that won't be today I think...
Btw - I was looking at FS state and I wonder would it be feasible
to make variant of this state using immutable map (probably some
scala one) to be able to do async checkpoints.
Then synchronous part would be essentially free - just taking the
state map and materializing it asynchronously.
Of course, that would work only for immutable state - but this is
often the case when writing in scala. WDYT?
thanks,
maciek
On 20/04/2016 16:28, Ufuk Celebi wrote:
Could be different things actually, including the parts of the
network
you mentioned.
1)
Regarding the TM config:
- It can help to increase the number of network buffers (you can go
ahead and give it 4 GB, e.g. 134217728 buffers a 32 KB)
- In general, you have way more memory available than you actually
give to Flink. I would increase the 20 GB heap size.
As a first step you could address these two points and re-run
your job.
2)
As a follow-up you could also work with the FileSystemStateBackend,
which keeps state in memory (on-heap) and writes checkpoints to
files.
This would help in checking how much RocksDB is slowing things
down.
I'm curious about the results. Do you think you will have time
to try this?
– Ufuk
On Wed, Apr 20, 2016 at 3:45 PM, Maciek Próchniak <m...@touk.pl>
<mailto:m...@touk.pl> wrote:
Hi,
I'm running my flink job on one rather large machine (20 cores
with
hyperthreading, 120GB RAM). Task manager has 20GB heap allocated.
It does more or less:
read csv from kafka -> keyBy one of the fields -> some custom
state
processing.
Kafka topic has 24 partitions, so my parallelism is also 24
After some tweaks and upgrading to 1.0.2-rc3 (as I use RocksDB
state
backend) I reached a point when throughput is ~120-150k/s.
One the same kafka and machine I reached > 500k/s with simple
filtering job,
so I wanted to see what's the bottleneck.
It turns out that quite often all of kafka threads are stuck
waiting for
buffer from pool:
"Thread-6695" #7466 daemon prio=5 os_prio=0 tid=0x00007f77fd80d000
nid=0x8118 in Object.wait() [0x00007f7ad54d9000]
java.lang.Thread.State: TIMED_WAITING (on object monitor)
at java.lang.Object.wait(Native Method)
at
org.apache.flink.runtime.io.network.buffer.LocalBufferPool.requestBuffer(LocalBufferPool.java:163)
- locked <0x00000002eade3890> (a java.util.ArrayDeque)
at
org.apache.flink.runtime.io.network.buffer.LocalBufferPool.requestBufferBlocking(LocalBufferPool.java:133)
at
org.apache.flink.runtime.io.network.api.writer.RecordWriter.emit(RecordWriter.java:92)
- locked <0x00000002eb73cbd0> (a
org.apache.flink.runtime.io.network.api.serialization.SpanningRecordSerializer)
at
org.apache.flink.streaming.runtime.io.StreamRecordWriter.emit(StreamRecordWriter.java:86)
at
org.apache.flink.streaming.runtime.io.RecordWriterOutput.collect(RecordWriterOutput.java:78)
at
org.apache.flink.streaming.runtime.io.RecordWriterOutput.collect(RecordWriterOutput.java:39)
at
org.apache.flink.streaming.api.operators.TimestampedCollector.collect(TimestampedCollector.java:51)
at
org.apache.flink.streaming.api.scala.DataStream$$anon$6$$anonfun$flatMap$1.apply(DataStream.scala:541)
at
org.apache.flink.streaming.api.scala.DataStream$$anon$6$$anonfun$flatMap$1.apply(DataStream.scala:541)
at
scala.collection.immutable.List.foreach(List.scala:381)
at
org.apache.flink.streaming.api.scala.DataStream$$anon$6.flatMap(DataStream.scala:541)
at
org.apache.flink.streaming.api.operators.StreamFlatMap.processElement(StreamFlatMap.java:48)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:309)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:297)
at
org.apache.flink.streaming.api.operators.StreamFilter.processElement(StreamFilter.java:38)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:309)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:297)
at
org.apache.flink.streaming.api.operators.StreamFilter.processElement(StreamFilter.java:38)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:309)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:297)
at
org.apache.flink.streaming.api.operators.StreamMap.processElement(StreamMap.java:39)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:309)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:297)
at
org.apache.flink.streaming.api.operators.StreamMap.processElement(StreamMap.java:39)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:309)
at
org.apache.flink.streaming.runtime.tasks.OperatorChain$ChainingOutput.collect(OperatorChain.java:297)
at
org.apache.flink.streaming.api.operators.StreamSource$ManualWatermarkContext.collect(StreamSource.java:318)
- locked <0x00000002eaf3eb50> (a java.lang.Object)
at
org.apache.flink.streaming.connectors.kafka.FlinkKafkaConsumer09$ConsumerThread.run(FlinkKafkaConsumer09.java:473)
- locked <0x00000002eaf3eb50> (a java.lang.Object)
This seems a bit weird for me, as most of state processing
threads are idle:
"My custom function -> (Sink: Unnamed, Map) (19/24)" #7353
daemon prio=5
os_prio=0 tid=0x00007f7a7400e000 nid=0x80a7 waiting on condition
[0x00007f7bee8ed000]
java.lang.Thread.State: TIMED_WAITING (parking)
at sun.misc.Unsafe.park(Native Method)
- parking to wait for <0x00000002eb840c38> (a
java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
at
java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:215)
at
java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(AbstractQueuedSynchronizer.java:2078)
at
java.util.concurrent.LinkedBlockingQueue.poll(LinkedBlockingQueue.java:467)
at
org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.getNextBufferOrEvent(SingleInputGate.java:415)
at
org.apache.flink.streaming.runtime.io.BarrierBuffer.getNextNonBlocked(BarrierBuffer.java:108)
at
org.apache.flink.streaming.runtime.io.StreamInputProcessor.processInput(StreamInputProcessor.java:175)
at
org.apache.flink.streaming.runtime.tasks.OneInputStreamTask.run(OneInputStreamTask.java:65)
at
org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:225)
at
org.apache.flink.runtime.taskmanager.Task.run(Task.java:559)
at java.lang.Thread.run(Thread.java:745)
I tried with using more network buffers, but I doesn't seem to
change
anything - and if I understand correctly
https://ci.apache.org/projects/flink/flink-docs-master/setup/config.html#configuring-the-network-buffers
I should not need more than 24^2 * 4 of them...
Does anybody encountered such problem? Or maybe it's just
normal for such
case...
thanks,
maciek
