rickyma commented on PR #1759:
URL: 
https://github.com/apache/incubator-uniffle/pull/1759#issuecomment-2147856361

   > And it seems that the larger the spill value is, the more off-heap memory 
will be used. That's why the executors are killed by yarn and i don't have time 
to find out more details currently.
   
   I can try to explain this to you. When the size of the spill is larger, it 
means that more data is accumulated at the client side each time, so the amount 
of data sent each time is also larger. Therefore, more server responses are 
received in the same time period. After the netty mode is enabled, this part of 
the responses will occupy off-heap memory, which means that the off-heap memory 
may be used more during this time. If the off-heap memory configured for your 
executor is not large enough, it may cause OOM. This is why I have been 
recommending you to use the `spark.rss.client.memory.spill.ratio` parameter. If 
you enable this parameter and set it to less than 1.0, it means that you only 
send part of the spill data each time, which makes the client's requests 
smoother overall. When the value of the ratio is reached, a part is sent, 
instead of waiting until the end to send all at once. If you set up an off-heap 
memory monitoring for the client side, you will find that if the rati
 o is enabled, the used off-heap memory curve will be smoother. If the ratio is 
1.0 (not enabled), the curve will be jagged, and the memory used at the peak 
may be very large, which may cause OOM.
   
   Below is a comparison graph for different memory spill ratios:
   
![output](https://github.com/apache/incubator-uniffle/assets/13834479/4fa8b7ba-6e52-4f90-bcef-c1b4258067c2)
   
   So again, `spark.rss.client.memory.spill.ratio` should help you mitigate 
executors being killed by yarn problems.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to