andygrove commented on PR #3536:
URL: 
https://github.com/apache/datafusion-comet/pull/3536#issuecomment-3910056037

   > > The changes in this PR are somehow causing large differences in memory 
usage compared to a recent build from the main branch.
   > > Metric   Baseline        DF52    Delta
   > > peak_JVMHeapMemory       6.38 GB 10.60 GB        +4.22 GB (+66%)
   > > peak_JVMOffHeapMemory    148.01 MB       147.28 MB       -0.73 MB (~0%)
   > > peak_OnHeapExecutionMemory       0 B     0 B     --
   > > peak_OffHeapExecutionMemory      14.20 GB        5.30 GB -8.90 GB (-63%)
   > > peak_OnHeapUnifiedMemory 74.86 MB        63.63 MB        -11.24 MB (-15%)
   > > peak_OffHeapUnifiedMemory        14.20 GB        5.30 GB -8.90 GB (-63%)
   > > DF52 shows a significant shift from off-heap execution memory to JVM 
heap. Peak off-heap execution dropped from 14.20 GB to 5.30 GB (-63%), while 
JVM heap rose from 6.38 GB to 10.60 GB (+66%). Net combined peak memory 
actually decreased (~20.6 GB to ~15.9 GB).
   > 
   > The total mem used seems to be less? +4 for Heap and -8 for on Heap, so -4 
total. The JVM growth looks weird as DF52 has very few JVM changes. Is the 
baseline for latest main?
   
   The baseline was a build from Friday.
   
   Claude is suggesting this as the root cause:
   
   "The memory shift is primarily driven by DataFusion 52's internal changes to 
how operators manage buffers (batch coalescing, expression evaluation), 
combined with these buffers likely not being tracked through Comet's 
CometUnifiedMemoryPool → JNI → Spark accounting path. The result is that native 
memory still exists but is "invisible" to Spark's off-heap accounting, while 
Spark's JVM-side operations expand to fill the perceived available space."
   
   
   


-- 
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