You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@spark.apache.org by Dhrubajyoti Hati <dh...@gmail.com> on 2019/07/29 06:22:11 UTC
Logistic Regression Iterations causing High GC in Spark 2.3
Hi,
We were running Logistic Regression in Spark 2.2.X and then we tried to see
how does it do in Spark 2.3.X. Now we are facing an issue while running a
Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In
the TreeAggregate method it takes a huge time due to very High GC Activity.
I have tuned the GC, created different sized clusters, higher spark
version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000
times of the processing time in avg for iterations.
The strange part is in *Spark 2.2 this doesn't happen at all*. Same code,
same cluster sizing, same data in both the cases.
I was wondering if someone can explain this behaviour and help me to
resolve this. How can the same code has so different behaviour in two Spark
version, especially the higher ones?
Here are the config which I used:
spark.serializer=org.apache.spark.serializer.KryoSerializer
#GC Tuning
spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal
-XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails
-XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy
-XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m
-XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
spark.executor.instances=20
spark.executor.cores=1
spark.executor.memory=9010m
Regards,
Dhrub
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Dhrubajyoti Hati <dh...@gmail.com>.
Actually the original data is around ~120 GB. If we provide higher memory
then we might require an even bigger cluster to finish training the whole
model within planned time. And this will affect the cost of operations.
Please correct me if I am wrong here.
Nevertheless, can you point out how much memory should be sufficient for
each executor. I already gave 9GB of Memory with 20 executors to process
the data.
On Mon, Jul 29, 2019 at 7:42 PM Sean Owen <sr...@gmail.com> wrote:
> Could be lots of things. Implementations change, caching may have
> changed, etc. The size of the input doesn't really directly translate
> to heap usage. Here you just need a bit more memory.
>
> On Mon, Jul 29, 2019 at 9:03 AM Dhrubajyoti Hati <dh...@gmail.com>
> wrote:
> >
> > Hi Sean,
> >
> > Yeah I checked the heap, its almost full. I checked the GC logs in the
> executors where I found that GC cycles are kicking in frequently. The
> Executors tab shows red in the "Total Time/GC Time".
> >
> > Also the data which I am dealing with is quite small(~4 GB) and the
> cluster is quite big for that high GC.
> >
> > But what's troubling me is this issue doesn't occur in Spark 2.2 at all.
> What could be the reason behind such a behaviour?
> >
> > Regards,
> > Dhrub
> >
> > On Mon, Jul 29, 2019 at 6:45 PM Sean Owen <sr...@gmail.com> wrote:
> >>
> >> -dev@
> >>
> >> Yep, high GC activity means '(almost) out of memory'. I don't see that
> >> you've checked heap usage - is it nearly full?
> >> The answer isn't tuning but more heap.
> >> (Sometimes with really big heaps the problem is big pauses, but that's
> >> not the case here.)
> >>
> >> On Mon, Jul 29, 2019 at 1:26 AM Dhrubajyoti Hati <dh...@gmail.com>
> wrote:
> >> >
> >> > Hi,
> >> >
> >> > We were running Logistic Regression in Spark 2.2.X and then we tried
> to see how does it do in Spark 2.3.X. Now we are facing an issue while
> running a Logistic Regression Model in Spark 2.3.X on top of
> Yarn(GCP-Dataproc). In the TreeAggregate method it takes a huge time due to
> very High GC Activity. I have tuned the GC, created different sized
> clusters, higher spark version(2.4.X), smaller data but nothing helps. The
> GC time is 100 - 1000 times of the processing time in avg for iterations.
> >> >
> >> > The strange part is in Spark 2.2 this doesn't happen at all. Same
> code, same cluster sizing, same data in both the cases.
> >> >
> >> > I was wondering if someone can explain this behaviour and help me to
> resolve this. How can the same code has so different behaviour in two Spark
> version, especially the higher ones?
> >> >
> >> > Here are the config which I used:
> >> >
> >> >
> >> > spark.serializer=org.apache.spark.serializer.KryoSerializer
> >> >
> >> > #GC Tuning
> >> >
> >> > spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal
> -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails
> -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy
> -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m
> -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
> >> >
> >> >
> >> > spark.executor.instances=20
> >> >
> >> > spark.executor.cores=1
> >> >
> >> > spark.executor.memory=9010m
> >> >
> >> >
> >> >
> >> > Regards,
> >> > Dhrub
> >> >
>
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Sean Owen <sr...@gmail.com>.
Could be lots of things. Implementations change, caching may have
changed, etc. The size of the input doesn't really directly translate
to heap usage. Here you just need a bit more memory.
On Mon, Jul 29, 2019 at 9:03 AM Dhrubajyoti Hati <dh...@gmail.com> wrote:
>
> Hi Sean,
>
> Yeah I checked the heap, its almost full. I checked the GC logs in the executors where I found that GC cycles are kicking in frequently. The Executors tab shows red in the "Total Time/GC Time".
>
> Also the data which I am dealing with is quite small(~4 GB) and the cluster is quite big for that high GC.
>
> But what's troubling me is this issue doesn't occur in Spark 2.2 at all. What could be the reason behind such a behaviour?
>
> Regards,
> Dhrub
>
> On Mon, Jul 29, 2019 at 6:45 PM Sean Owen <sr...@gmail.com> wrote:
>>
>> -dev@
>>
>> Yep, high GC activity means '(almost) out of memory'. I don't see that
>> you've checked heap usage - is it nearly full?
>> The answer isn't tuning but more heap.
>> (Sometimes with really big heaps the problem is big pauses, but that's
>> not the case here.)
>>
>> On Mon, Jul 29, 2019 at 1:26 AM Dhrubajyoti Hati <dh...@gmail.com> wrote:
>> >
>> > Hi,
>> >
>> > We were running Logistic Regression in Spark 2.2.X and then we tried to see how does it do in Spark 2.3.X. Now we are facing an issue while running a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In the TreeAggregate method it takes a huge time due to very High GC Activity. I have tuned the GC, created different sized clusters, higher spark version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000 times of the processing time in avg for iterations.
>> >
>> > The strange part is in Spark 2.2 this doesn't happen at all. Same code, same cluster sizing, same data in both the cases.
>> >
>> > I was wondering if someone can explain this behaviour and help me to resolve this. How can the same code has so different behaviour in two Spark version, especially the higher ones?
>> >
>> > Here are the config which I used:
>> >
>> >
>> > spark.serializer=org.apache.spark.serializer.KryoSerializer
>> >
>> > #GC Tuning
>> >
>> > spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
>> >
>> >
>> > spark.executor.instances=20
>> >
>> > spark.executor.cores=1
>> >
>> > spark.executor.memory=9010m
>> >
>> >
>> >
>> > Regards,
>> > Dhrub
>> >
---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Dhrubajyoti Hati <dh...@gmail.com>.
Hi Sean,
Yeah I checked the heap, its almost full. I checked the GC logs in the
executors where I found that GC cycles are kicking in frequently. The
Executors tab shows red in the "Total Time/GC Time".
Also the data which I am dealing with is quite small(~4 GB) and the cluster
is quite big for that high GC.
But what's troubling me is this issue doesn't occur in Spark 2.2 at all.
What could be the reason behind such a behaviour?
Regards,
Dhrub
On Mon, Jul 29, 2019 at 6:45 PM Sean Owen <sr...@gmail.com> wrote:
> -dev@
>
> Yep, high GC activity means '(almost) out of memory'. I don't see that
> you've checked heap usage - is it nearly full?
> The answer isn't tuning but more heap.
> (Sometimes with really big heaps the problem is big pauses, but that's
> not the case here.)
>
> On Mon, Jul 29, 2019 at 1:26 AM Dhrubajyoti Hati <dh...@gmail.com>
> wrote:
> >
> > Hi,
> >
> > We were running Logistic Regression in Spark 2.2.X and then we tried to
> see how does it do in Spark 2.3.X. Now we are facing an issue while running
> a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In
> the TreeAggregate method it takes a huge time due to very High GC Activity.
> I have tuned the GC, created different sized clusters, higher spark
> version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000
> times of the processing time in avg for iterations.
> >
> > The strange part is in Spark 2.2 this doesn't happen at all. Same code,
> same cluster sizing, same data in both the cases.
> >
> > I was wondering if someone can explain this behaviour and help me to
> resolve this. How can the same code has so different behaviour in two Spark
> version, especially the higher ones?
> >
> > Here are the config which I used:
> >
> >
> > spark.serializer=org.apache.spark.serializer.KryoSerializer
> >
> > #GC Tuning
> >
> > spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal
> -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails
> -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy
> -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m
> -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
> >
> >
> > spark.executor.instances=20
> >
> > spark.executor.cores=1
> >
> > spark.executor.memory=9010m
> >
> >
> >
> > Regards,
> > Dhrub
> >
>
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Sean Owen <sr...@gmail.com>.
-dev@
Yep, high GC activity means '(almost) out of memory'. I don't see that
you've checked heap usage - is it nearly full?
The answer isn't tuning but more heap.
(Sometimes with really big heaps the problem is big pauses, but that's
not the case here.)
On Mon, Jul 29, 2019 at 1:26 AM Dhrubajyoti Hati <dh...@gmail.com> wrote:
>
> Hi,
>
> We were running Logistic Regression in Spark 2.2.X and then we tried to see how does it do in Spark 2.3.X. Now we are facing an issue while running a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In the TreeAggregate method it takes a huge time due to very High GC Activity. I have tuned the GC, created different sized clusters, higher spark version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000 times of the processing time in avg for iterations.
>
> The strange part is in Spark 2.2 this doesn't happen at all. Same code, same cluster sizing, same data in both the cases.
>
> I was wondering if someone can explain this behaviour and help me to resolve this. How can the same code has so different behaviour in two Spark version, especially the higher ones?
>
> Here are the config which I used:
>
>
> spark.serializer=org.apache.spark.serializer.KryoSerializer
>
> #GC Tuning
>
> spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
>
>
> spark.executor.instances=20
>
> spark.executor.cores=1
>
> spark.executor.memory=9010m
>
>
>
> Regards,
> Dhrub
>
---------------------------------------------------------------------
To unsubscribe e-mail: user-unsubscribe@spark.apache.org
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Dhrubajyoti Hati <dh...@gmail.com>.
Actually I didn't have any of the GC tuning in the beginning and then
adding them also didn't made any difference. As mentioned earlier I tried
low number executors of higher configuration and vice versa. Nothing helps.
About the code its simple logistic regression nothing with explicit
broadcast or anything. The data is stored as parquet files in GCS bucket.
*The question is how does it work fine with Spark 2.2 and has this issue
with Spark 2.3 or higher version. As I mentioned before same code, same
cluster configuration and size, same data in both the cases.*
Regards,
Dhrub
On Mon, Jul 29, 2019 at 12:37 PM Jörn Franke <jo...@gmail.com> wrote:
> I would remove the all GC tuning and add it later once you found the
> underlying root cause. Usually more GC means you need to provide more
> memory, because something has changed (your application, spark Version etc.)
>
> We don’t have your full code to give exact advise, but you may want to
> rethink the one code / executor approach and have less executors but more
> cores / executor. That sometimes can lead to more heap usage (especially if
> you broadcast). Keep in mind that if you use more cores/executor it usually
> also requires more memory for the executor, but less executors. Similarly
> the executor instances might be too many and they may not have enough heap.
> You can also increase the memory of the executor.
>
> Am 29.07.2019 um 08:22 schrieb Dhrubajyoti Hati <dh...@gmail.com>:
>
> Hi,
>
> We were running Logistic Regression in Spark 2.2.X and then we tried to
> see how does it do in Spark 2.3.X. Now we are facing an issue while running
> a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In
> the TreeAggregate method it takes a huge time due to very High GC Activity.
> I have tuned the GC, created different sized clusters, higher spark
> version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000
> times of the processing time in avg for iterations.
>
> The strange part is in *Spark 2.2 this doesn't happen at all*. Same code,
> same cluster sizing, same data in both the cases.
>
> I was wondering if someone can explain this behaviour and help me to
> resolve this. How can the same code has so different behaviour in two Spark
> version, especially the higher ones?
>
> Here are the config which I used:
>
>
> spark.serializer=org.apache.spark.serializer.KryoSerializer
>
> #GC Tuning
>
> spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal
> -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails
> -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy
> -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m
> -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
>
>
> spark.executor.instances=20
>
> spark.executor.cores=1
>
> spark.executor.memory=9010m
>
>
> Regards,
> Dhrub
>
>
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Dhrubajyoti Hati <dh...@gmail.com>.
Actually I didn't have any of the GC tuning in the beginning and then
adding them also didn't made any difference. As mentioned earlier I tried
low number executors of higher configuration and vice versa. Nothing helps.
About the code its simple logistic regression nothing with explicit
broadcast or anything. The data is stored as parquet files in GCS bucket.
*The question is how does it work fine with Spark 2.2 and has this issue
with Spark 2.3 or higher version. As I mentioned before same code, same
cluster configuration and size, same data in both the cases.*
Regards,
Dhrub
On Mon, Jul 29, 2019 at 12:37 PM Jörn Franke <jo...@gmail.com> wrote:
> I would remove the all GC tuning and add it later once you found the
> underlying root cause. Usually more GC means you need to provide more
> memory, because something has changed (your application, spark Version etc.)
>
> We don’t have your full code to give exact advise, but you may want to
> rethink the one code / executor approach and have less executors but more
> cores / executor. That sometimes can lead to more heap usage (especially if
> you broadcast). Keep in mind that if you use more cores/executor it usually
> also requires more memory for the executor, but less executors. Similarly
> the executor instances might be too many and they may not have enough heap.
> You can also increase the memory of the executor.
>
> Am 29.07.2019 um 08:22 schrieb Dhrubajyoti Hati <dh...@gmail.com>:
>
> Hi,
>
> We were running Logistic Regression in Spark 2.2.X and then we tried to
> see how does it do in Spark 2.3.X. Now we are facing an issue while running
> a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In
> the TreeAggregate method it takes a huge time due to very High GC Activity.
> I have tuned the GC, created different sized clusters, higher spark
> version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000
> times of the processing time in avg for iterations.
>
> The strange part is in *Spark 2.2 this doesn't happen at all*. Same code,
> same cluster sizing, same data in both the cases.
>
> I was wondering if someone can explain this behaviour and help me to
> resolve this. How can the same code has so different behaviour in two Spark
> version, especially the higher ones?
>
> Here are the config which I used:
>
>
> spark.serializer=org.apache.spark.serializer.KryoSerializer
>
> #GC Tuning
>
> spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal
> -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails
> -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy
> -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m
> -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
>
>
> spark.executor.instances=20
>
> spark.executor.cores=1
>
> spark.executor.memory=9010m
>
>
> Regards,
> Dhrub
>
>
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Jörn Franke <jo...@gmail.com>.
I would remove the all GC tuning and add it later once you found the underlying root cause. Usually more GC means you need to provide more memory, because something has changed (your application, spark Version etc.)
We don’t have your full code to give exact advise, but you may want to rethink the one code / executor approach and have less executors but more cores / executor. That sometimes can lead to more heap usage (especially if you broadcast). Keep in mind that if you use more cores/executor it usually also requires more memory for the executor, but less executors. Similarly the executor instances might be too many and they may not have enough heap.
You can also increase the memory of the executor.
> Am 29.07.2019 um 08:22 schrieb Dhrubajyoti Hati <dh...@gmail.com>:
>
> Hi,
>
> We were running Logistic Regression in Spark 2.2.X and then we tried to see how does it do in Spark 2.3.X. Now we are facing an issue while running a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In the TreeAggregate method it takes a huge time due to very High GC Activity. I have tuned the GC, created different sized clusters, higher spark version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000 times of the processing time in avg for iterations.
>
> The strange part is in Spark 2.2 this doesn't happen at all. Same code, same cluster sizing, same data in both the cases.
>
> I was wondering if someone can explain this behaviour and help me to resolve this. How can the same code has so different behaviour in two Spark version, especially the higher ones?
>
> Here are the config which I used:
>
> spark.serializer=org.apache.spark.serializer.KryoSerializer
> #GC Tuning
> spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
>
> spark.executor.instances=20
> spark.executor.cores=1
> spark.executor.memory=9010m
>
>
> Regards,
> Dhrub
>
Re: Logistic Regression Iterations causing High GC in Spark 2.3
Posted by Jörn Franke <jo...@gmail.com>.
I would remove the all GC tuning and add it later once you found the underlying root cause. Usually more GC means you need to provide more memory, because something has changed (your application, spark Version etc.)
We don’t have your full code to give exact advise, but you may want to rethink the one code / executor approach and have less executors but more cores / executor. That sometimes can lead to more heap usage (especially if you broadcast). Keep in mind that if you use more cores/executor it usually also requires more memory for the executor, but less executors. Similarly the executor instances might be too many and they may not have enough heap.
You can also increase the memory of the executor.
> Am 29.07.2019 um 08:22 schrieb Dhrubajyoti Hati <dh...@gmail.com>:
>
> Hi,
>
> We were running Logistic Regression in Spark 2.2.X and then we tried to see how does it do in Spark 2.3.X. Now we are facing an issue while running a Logistic Regression Model in Spark 2.3.X on top of Yarn(GCP-Dataproc). In the TreeAggregate method it takes a huge time due to very High GC Activity. I have tuned the GC, created different sized clusters, higher spark version(2.4.X), smaller data but nothing helps. The GC time is 100 - 1000 times of the processing time in avg for iterations.
>
> The strange part is in Spark 2.2 this doesn't happen at all. Same code, same cluster sizing, same data in both the cases.
>
> I was wondering if someone can explain this behaviour and help me to resolve this. How can the same code has so different behaviour in two Spark version, especially the higher ones?
>
> Here are the config which I used:
>
> spark.serializer=org.apache.spark.serializer.KryoSerializer
> #GC Tuning
> spark.executor.extraJavaOptions= -XX:+UseG1GC -XX:+PrintFlagsFinal -XX:+PrintReferenceGC -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps -XX:+PrintAdaptiveSizePolicy -XX:+UnlockDiagnosticVMOptions -XX:+G1SummarizeConcMark -Xms9000m -XX:ParallelGCThreads=20 -XX:ConcGCThreads=5
>
> spark.executor.instances=20
> spark.executor.cores=1
> spark.executor.memory=9010m
>
>
> Regards,
> Dhrub
>