From 5899c891cb08d94bd16a2af40019ee61064ff86b Mon Sep 17 00:00:00 2001 From: Alessandro Bellina Date: Tue, 14 May 2024 11:58:15 -0700 Subject: [PATCH 1/2] Disabling the cuDF default pinned pool for 24.06 Signed-off-by: Alessandro Bellina --- .../main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala | 2 ++ 1 file changed, 2 insertions(+) diff --git a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala index 94390100d2e..345aed59f03 100644 --- a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala +++ b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala @@ -447,6 +447,8 @@ object GpuDeviceManager extends Logging { } else { (conf.pinnedPoolSize, -1L) } + // disable the cuDF provided default pinned pool for now + PinnedMemoryPool.configureDefaultCudfPinnedPoolSize(0L) if (!PinnedMemoryPool.isInitialized && pinnedSize > 0) { logInfo(s"Initializing pinned memory pool (${pinnedSize / 1024 / 1024.0} MiB)") PinnedMemoryPool.initialize(pinnedSize, gpuId, setCuioDefaultResource) From e437abdb6b36babef2ae2cf6c5c25e922b07fc98 Mon Sep 17 00:00:00 2001 From: Alessandro Bellina Date: Wed, 22 May 2024 12:50:05 -0700 Subject: [PATCH 2/2] Add a warning in case we can't configure the cuDF default pool --- .../scala/com/nvidia/spark/rapids/GpuDeviceManager.scala | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala index 345aed59f03..2cfce60c4a5 100644 --- a/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala +++ b/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala @@ -448,7 +448,11 @@ object GpuDeviceManager extends Logging { (conf.pinnedPoolSize, -1L) } // disable the cuDF provided default pinned pool for now - PinnedMemoryPool.configureDefaultCudfPinnedPoolSize(0L) + if (!PinnedMemoryPool.configureDefaultCudfPinnedPoolSize(0L)) { + // This is OK in tests because they don't unload/reload our shared + // library, and in prod it would be nice to know about it. + logWarning("The default cuDF host pool was already configured") + } if (!PinnedMemoryPool.isInitialized && pinnedSize > 0) { logInfo(s"Initializing pinned memory pool (${pinnedSize / 1024 / 1024.0} MiB)") PinnedMemoryPool.initialize(pinnedSize, gpuId, setCuioDefaultResource)