Re: [EXTERNAL] Re: Re: Re: Stage level scheduling - lower the number of executors when using GPUs
We're actually running on on-prem Kubernetes with a custom-built build Spark image, with altered entrypoint.sh and other "low-level" scripts and configs, but I don't think this is a good direction to solve this specific issue. Shay From: Artemis User Sent: Thursday, November 3, 2022 8:35 PM To: user@spark.apache.org Subject: [EXTERNAL] Re: Re: Re: Stage level scheduling - lower the number of executors when using GPUs ATTENTION: This email originated from outside of GM. Now I see what you want to do. If you have access to the cluster configuration files, you can modify the spark-env.sh file on the worker nodes to specify exactly which node you'd like to link with GPU cores and which one not. This would allow only those nodes configured with GPU-resources getting scheduled/acquired for your GPU tasks (see Rapids user guide at https://nvidia.github.io/spark-rapids/docs/get-started/getting-started-on-prem.html). We are using Rapids in our on-prem Spark environment with complete control of OS, file and network systems, containers and even hardware/GPU settings. I guess you are using one of the cloud services so I am not sure if you have access to the low-level cluster config on EMR or GCP, which gave you a cookie-cutter type of cluster settings with limited configurability. But under the hood, I believe they do use Nvidia Rapids which currently is the only option for GPU acceleration in Spark (Spark 3.x.x distribution package doesn't include Rapids or any GPU integration libs). So you may want to dive into the Rapids instructions for more configuration and usage info (it does provide detailed instructions on how to run Rapids on EMR, Databricks and GCP). On 11/3/22 12:10 PM, Shay Elbaz wrote: Thanks again Artemis, I really appreciate it. I have watched the video but did not find an answer. Please bear with me just one more iteration Maybe I'll be more specific: Suppose I start the application with maxExecutors=500, executors.cores=2, because that's the amount of resources needed for the ETL part. But for the DL part I only need 20 GPUs. SLS API only allows to set the resources per executor/task, so Spark would (try to) allocate up to 500 GPUs, assuming I configure the profile with 1 GPU per executor. So, the question is how do I limit the stage resources to 20 GPUs total? Thanks again, Shay From: Artemis User <mailto:arte...@dtechspace.com> Sent: Thursday, November 3, 2022 5:23 PM To: user@spark.apache.org<mailto:user@spark.apache.org> <mailto:user@spark.apache.org> Subject: [EXTERNAL] Re: Re: Stage level scheduling - lower the number of executors when using GPUs ATTENTION: This email originated from outside of GM. Shay, You may find this video helpful (with some API code samples that you are looking for). https://www.youtube.com/watch?v=JNQu-226wUc=171s. The issue here isn't how to limit the number of executors but to request for the right GPU-enabled executors dynamically. Those executors used in pre-GPU stages should be returned back to resource managers with dynamic resource allocation enabled (and with the right DRA policies). Hope this helps.. Unfortunately there isn't a lot of detailed docs for this topic since GPU acceleration is kind of new in Spark (not straightforward like in TF). I wish the Spark doc team could provide more details in the next release... On 11/3/22 2:37 AM, Shay Elbaz wrote: Thanks Artemis. We are not using Rapids, but rather using GPUs through the Stage Level Scheduling feature with ResourceProfile. In Kubernetes you have to turn on shuffle tracking for dynamic allocation, anyhow. The question is how we can limit the number of executors when building a new ResourceProfile, directly (API) or indirectly (some advanced workaround). Thanks, Shay From: Artemis User <mailto:arte...@dtechspace.com> Sent: Thursday, November 3, 2022 1:16 AM To: user@spark.apache.org<mailto:user@spark.apache.org> <mailto:user@spark.apache.org> Subject: [EXTERNAL] Re: Stage level scheduling - lower the number of executors when using GPUs ATTENTION: This email originated from outside of GM. Are you using Rapids for GPU support in Spark? Couple of options you may want to try: 1. In addition to dynamic allocation turned on, you may also need to turn on external shuffling service. 2. Sounds like you are using Kubernetes. In that case, you may also need to turn on shuffle tracking. 3. The "stages" are controlled by the APIs. The APIs for dynamic resource request (change of stage) do exist, but only for RDDs (e.g. TaskResourceRequest and ExecutorResourceRequest). On 11/2/22 11:30 AM, Shay Elbaz wrote: Hi, Our typical applications need less executors for a GPU stage than for a CPU stage. We are using dynamic allocation with stage level scheduling, and Spark tries
Re: [EXTERNAL] Re: Re: Re: Stage level scheduling - lower the number of executors when using GPUs
So I'm not sure I completely follow. Are you asking for a way to change the
limit without having to do the repartition? And your DL software doesn't care
if you got say 30 executors instead of 20? Normally I would expect the number
fo partitions at that point to be 200 (or whatever you set for your shuffle
partitions) unless you are using AQE coalescing partitions functionality and
then it could change. Are you using the latter?
> Normally I try to aim for anything between 30s-5m per task (failure-wise),
> depending on the cluster, its stability, etc. But in this case, individual
> tasks can take 30-60 minutes, if not much more. Any failure during this long
> time is pretty expensive.
Are you saying when you manually do the repartition your DL tasks take 30-60
minutes? so again you want like AQE coalesce partitions to kick in to attempt
to pick partition sizes for your?
Tom
On Thursday, November 3, 2022 at 03:18:07 PM CDT, Shay Elbaz
wrote:
#yiv4404278030 P {margin-top:0;margin-bottom:0;}This is exactly what we ended
up doing! The only drawback I saw with this approach is that the GPU tasks get
pretty big (in terms of data and compute time), and task failures become
expansive. That's why I reached out to the mailing list in the first place
Normally I try to aim for anything between 30s-5m per task (failure-wise),
depending on the cluster, its stability, etc. But in this case, individual
tasks can take 30-60 minutes, if not much more. Any failure during this long
time is pretty expensive.
ShayFrom: Tom Graves
Sent: Thursday, November 3, 2022 7:56 PM
To: Artemis User ; user@spark.apache.org
; Shay Elbaz
Subject: [EXTERNAL] Re: Re: Re: Stage level scheduling - lower the number of
executors when using GPUs
|
ATTENTION: This email originated from outside of GM.
|
Stage level scheduling does not allow you to change configs right now. This is
something we thought about as follow on but have never implemented. How many
tasks on the DL stage are you running? The typical case is run some etl lots
of tasks... do mapPartitions and then run your DL stuff, before that
mapPartitions you could do a repartition if necessary to get to exactly the
number of tasks you want (20). That way even if maxExecutors=500 you will only
ever need 20 or whatever you repartition to and spark isn't going to ask for
more then that.
Tom
On Thursday, November 3, 2022 at 11:10:31 AM CDT, Shay Elbaz
wrote:
Thanks again Artemis, I really appreciate it.
I have watched the video but did not find an answer.
Please bear with me just one more iteration
Maybe I'll be more specific:Suppose I start the application with
maxExecutors=500, executors.cores=2, because that's the amount of resources
needed for the ETL part. But for the DL part I only need 20 GPUs. SLS API only
allows to set the resources per executor/task, so Spark would (try to) allocate
up to 500 GPUs, assuming I configure the profile with 1 GPU per executor. So,
the question is how do I limit the stage resources to 20 GPUs total?
Thanks again,Shay
From: Artemis User
Sent: Thursday, November 3, 2022 5:23 PM
To: user@spark.apache.org
Subject: [EXTERNAL] Re: Re: Stage level scheduling - lower the number of
executors when using GPUs
|
ATTENTION: This email originated from outside of GM.
|
Shay, You may find this video helpful (with some API code samples that you
are looking for). https://www.youtube.com/watch?v=JNQu-226wUc=171s. The
issue here isn't how to limit the number of executors but to request for the
right GPU-enabled executors dynamically. Those executors used in pre-GPU
stages should be returned back to resource managers with dynamic resource
allocation enabled (and with the right DRA policies). Hope this helps..
Unfortunately there isn't a lot of detailed docs for this topic since GPU
acceleration is kind of new in Spark (not straightforward like in TF). I wish
the Spark doc team could provide more details in the next release...
On 11/3/22 2:37 AM, Shay Elbaz wrote:
Thanks Artemis. We are not using Rapids,
but rather using GPUs through the Stage Level Scheduling feature with
ResourceProfile. In Kubernetes you have to turn on shuffle tracking for dynamic
allocation, anyhow.The question is how we can limit thenumber of executors when
building a new ResourceProfile, directly (API) or indirectly (some advanced
workaround).
Thanks,Shay
From: Artemis User
Sent: Thursday, November 3, 2022 1:16 AM
To: user@spark.apache.org
Subject: [EXTERNAL] Re: Stage level scheduling - lower the number of executors
when using GPUs
|
ATTENTION: This email originated from outside of GM.
|
Are you using Rapids for GPU support in Spark? Couple of options you may
want to try:
- In addition to dynamic allocation turned on, you may also need to turn on
external shuffling service.
- Sounds like you are using Kubernetes. In that case, you may also need to
turn on shuf
Re: [EXTERNAL] Re: Re: Re: Stage level scheduling - lower the number of executors when using GPUs
This is exactly what we ended up doing! The only drawback I saw with this approach is that the GPU tasks get pretty big (in terms of data and compute time), and task failures become expansive. That's why I reached out to the mailing list in the first place Normally I try to aim for anything between 30s-5m per task (failure-wise), depending on the cluster, its stability, etc. But in this case, individual tasks can take 30-60 minutes, if not much more. Any failure during this long time is pretty expensive. Shay From: Tom Graves Sent: Thursday, November 3, 2022 7:56 PM To: Artemis User ; user@spark.apache.org ; Shay Elbaz Subject: [EXTERNAL] Re: Re: Re: Stage level scheduling - lower the number of executors when using GPUs ATTENTION: This email originated from outside of GM. Stage level scheduling does not allow you to change configs right now. This is something we thought about as follow on but have never implemented. How many tasks on the DL stage are you running? The typical case is run some etl lots of tasks... do mapPartitions and then run your DL stuff, before that mapPartitions you could do a repartition if necessary to get to exactly the number of tasks you want (20). That way even if maxExecutors=500 you will only ever need 20 or whatever you repartition to and spark isn't going to ask for more then that. Tom On Thursday, November 3, 2022 at 11:10:31 AM CDT, Shay Elbaz wrote: Thanks again Artemis, I really appreciate it. I have watched the video but did not find an answer. Please bear with me just one more iteration Maybe I'll be more specific: Suppose I start the application with maxExecutors=500, executors.cores=2, because that's the amount of resources needed for the ETL part. But for the DL part I only need 20 GPUs. SLS API only allows to set the resources per executor/task, so Spark would (try to) allocate up to 500 GPUs, assuming I configure the profile with 1 GPU per executor. So, the question is how do I limit the stage resources to 20 GPUs total? Thanks again, Shay From: Artemis User Sent: Thursday, November 3, 2022 5:23 PM To: user@spark.apache.org Subject: [EXTERNAL] Re: Re: Stage level scheduling - lower the number of executors when using GPUs ATTENTION: This email originated from outside of GM. Shay, You may find this video helpful (with some API code samples that you are looking for). https://www.youtube.com/watch?v=JNQu-226wUc=171s. The issue here isn't how to limit the number of executors but to request for the right GPU-enabled executors dynamically. Those executors used in pre-GPU stages should be returned back to resource managers with dynamic resource allocation enabled (and with the right DRA policies). Hope this helps.. Unfortunately there isn't a lot of detailed docs for this topic since GPU acceleration is kind of new in Spark (not straightforward like in TF). I wish the Spark doc team could provide more details in the next release... On 11/3/22 2:37 AM, Shay Elbaz wrote: Thanks Artemis. We are not using Rapids, but rather using GPUs through the Stage Level Scheduling feature with ResourceProfile. In Kubernetes you have to turn on shuffle tracking for dynamic allocation, anyhow. The question is how we can limit the number of executors when building a new ResourceProfile, directly (API) or indirectly (some advanced workaround). Thanks, Shay From: Artemis User <mailto:arte...@dtechspace.com> Sent: Thursday, November 3, 2022 1:16 AM To: user@spark.apache.org<mailto:user@spark.apache.org> <mailto:user@spark.apache.org> Subject: [EXTERNAL] Re: Stage level scheduling - lower the number of executors when using GPUs ATTENTION: This email originated from outside of GM. Are you using Rapids for GPU support in Spark? Couple of options you may want to try: 1. In addition to dynamic allocation turned on, you may also need to turn on external shuffling service. 2. Sounds like you are using Kubernetes. In that case, you may also need to turn on shuffle tracking. 3. The "stages" are controlled by the APIs. The APIs for dynamic resource request (change of stage) do exist, but only for RDDs (e.g. TaskResourceRequest and ExecutorResourceRequest). On 11/2/22 11:30 AM, Shay Elbaz wrote: Hi, Our typical applications need less executors for a GPU stage than for a CPU stage. We are using dynamic allocation with stage level scheduling, and Spark tries to maximize the number of executors also during the GPU stage, causing a bit of resources chaos in the cluster. This forces us to use a lower value for 'maxExecutors' in the first place, at the cost of the CPU stages performance. Or try to solve this in the Kubernets scheduler level, which is not straightforward and doesn't feel like the right way to go. Is there a way to effectively use less executors in Stage Lev
