Post

Replies

Boosts

Views

Activity

Residency set memory not freed if process performs no GPU operation
Feedback report: FB23959296 If a process creates a residency set, calls requestResidency, endResidency, and then releases the residency set without ever having done any GPU operations, the memory from the residency set is not freed. Workaround: if the application runs any GPU operation (even an operation not involving the residency set) at any point in its lifecycle (before/while/after creating/releasing the residency set), the memory is freed properly. This was observed in the context of an application that makes an AI model resident in GPU-accessible memory. If the user unloads the model without running any prompts, the memory is not freed. The model occupies ~16GB of RAM so a lot of memory is being leaked. Reproduction: Store the repro.m and workaround.m files from below Run the following commands (repro.m demonstrates the bug; workaround.m demonstrates the workaround): $ clang -framework Foundation -framework Metal -o repro repro.m $ ./repro Footprint at start: 0.00 GB Footprint after buffer allocation: 4.30 GB Footprint 5s after teardown: 4.30 GB $ clang -framework Foundation -framework Metal -o workaround workaround.m $ ./workaround Footprint at start: 0.00 GB Footprint after buffer allocation: 4.37 GB Footprint 5s after teardown: 0.01 GB Expected behavior: Footprint 5s after teardown should be ~0 GB, i.e., the memory is freed. Observed behavior: Footprint 5s after teardown is 4.30 GB, i.e., the memory is not freed. Versions: XCode: 26.6 (17F113) Clang: 21.0.0 (clang-2100.1.1.101, arm64-apple-darwin25.5.0) macOS: 26.5.2 (25F84) Files: repro.m: // Build: clang -framework Foundation -framework Metal -o repro repro.m #import <Metal/Metal.h> #include <mach/mach.h> // Returns the physical memory footprint of the process. static double footprint_gb(void) { task_vm_info_data_t info; mach_msg_type_number_t n = TASK_VM_INFO_COUNT; task_info(mach_task_self(), TASK_VM_INFO, (task_info_t)&info, &n); return (double)info.phys_footprint / 1e9; } int main(int argc, char ** argv) { printf("Footprint at start: %5.2f GB\n", footprint_gb()); @autoreleasepool { id<MTLDevice> dev = MTLCreateSystemDefaultDevice(); // Allocate ~4GB of memory. const size_t size = 4ULL << 30; id<MTLBuffer> buf = [dev newBufferWithLength:size options:MTLResourceStorageModeShared]; memset(buf.contents, 0xab, size); // fault the pages in printf("Footprint after buffer allocation: %5.2f GB\n", footprint_gb()); MTLResidencySetDescriptor * desc = [[MTLResidencySetDescriptor alloc] init]; id<MTLResidencySet> rset = [dev newResidencySetWithDescriptor:desc error:nil]; [desc release]; [rset addAllocation:buf]; [rset commit]; [rset requestResidency]; [rset endResidency]; [rset removeAllAllocations]; [rset commit]; [rset release]; [buf release]; [dev release]; } sleep(5); printf("Footprint 5s after teardown: %5.2f GB\n", footprint_gb()); return 0; } workaround.m: // Build: clang -framework Foundation -framework Metal -o workaround workaround.m #import <Metal/Metal.h> #include <mach/mach.h> // Returns the physical memory footprint of the process. static double footprint_gb(void) { task_vm_info_data_t info; mach_msg_type_number_t n = TASK_VM_INFO_COUNT; task_info(mach_task_self(), TASK_VM_INFO, (task_info_t)&info, &n); return (double)info.phys_footprint / 1e9; } // Performing any work on the GPU ensures the memory from the residency set will be released. static void do_dummy_work(id<MTLDevice> dev, id<MTLCommandQueue> queue) { @autoreleasepool { id<MTLBuffer> tmp = [dev newBufferWithLength:1 options:MTLResourceStorageModeShared]; id<MTLCommandBuffer> cb = [queue commandBuffer]; id<MTLBlitCommandEncoder> enc = [cb blitCommandEncoder]; [enc fillBuffer:tmp range:NSMakeRange(0, 1) value:0]; [enc endEncoding]; [cb commit]; [tmp release]; } } int main(int argc, char ** argv) { printf("Footprint at start: %5.2f GB\n", footprint_gb()); @autoreleasepool { id<MTLDevice> dev = MTLCreateSystemDefaultDevice(); id<MTLCommandQueue> queue = [dev newCommandQueue]; // Workaround that ensures the memory will be released. // It also works if we call this after the residency set release or at any point in between. do_dummy_work(dev, queue); // Allocate ~4GB of memory. const size_t size = 4ULL << 30; id<MTLBuffer> buf = [dev newBufferWithLength:size options:MTLResourceStorageModeShared]; memset(buf.contents, 0xab, size); // fault the pages in printf("Footprint after buffer allocation: %5.2f GB\n", footprint_gb()); MTLResidencySetDescriptor * desc = [[MTLResidencySetDescriptor alloc] init]; id<MTLResidencySet> rset = [dev newResidencySetWithDescriptor:desc error:nil]; [desc release]; [rset addAllocation:buf]; [rset commit]; [rset requestResidency]; [rset endResidency]; [rset removeAllAllocations]; [rset commit]; [rset release]; [buf release]; [queue release]; [dev release]; } sleep(5); printf("Footprint 5s after teardown: %5.2f GB\n", footprint_gb()); return 0; }
0
0
38
9h
Residency set memory not freed if process performs no GPU operation
Feedback report: FB23959296 If a process creates a residency set, calls requestResidency, endResidency, and then releases the residency set without ever having done any GPU operations, the memory from the residency set is not freed. Workaround: if the application runs any GPU operation (even an operation not involving the residency set) at any point in its lifecycle (before/while/after creating/releasing the residency set), the memory is freed properly. This was observed in the context of an application that makes an AI model resident in GPU-accessible memory. If the user unloads the model without running any prompts, the memory is not freed. The model occupies ~16GB of RAM so a lot of memory is being leaked. Reproduction: Store the repro.m and workaround.m files from below Run the following commands (repro.m demonstrates the bug; workaround.m demonstrates the workaround): $ clang -framework Foundation -framework Metal -o repro repro.m $ ./repro Footprint at start: 0.00 GB Footprint after buffer allocation: 4.30 GB Footprint 5s after teardown: 4.30 GB $ clang -framework Foundation -framework Metal -o workaround workaround.m $ ./workaround Footprint at start: 0.00 GB Footprint after buffer allocation: 4.37 GB Footprint 5s after teardown: 0.01 GB Expected behavior: Footprint 5s after teardown should be ~0 GB, i.e., the memory is freed. Observed behavior: Footprint 5s after teardown is 4.30 GB, i.e., the memory is not freed. Versions: XCode: 26.6 (17F113) Clang: 21.0.0 (clang-2100.1.1.101, arm64-apple-darwin25.5.0) macOS: 26.5.2 (25F84) Files: repro.m: // Build: clang -framework Foundation -framework Metal -o repro repro.m #import <Metal/Metal.h> #include <mach/mach.h> // Returns the physical memory footprint of the process. static double footprint_gb(void) { task_vm_info_data_t info; mach_msg_type_number_t n = TASK_VM_INFO_COUNT; task_info(mach_task_self(), TASK_VM_INFO, (task_info_t)&info, &n); return (double)info.phys_footprint / 1e9; } int main(int argc, char ** argv) { printf("Footprint at start: %5.2f GB\n", footprint_gb()); @autoreleasepool { id<MTLDevice> dev = MTLCreateSystemDefaultDevice(); // Allocate ~4GB of memory. const size_t size = 4ULL << 30; id<MTLBuffer> buf = [dev newBufferWithLength:size options:MTLResourceStorageModeShared]; memset(buf.contents, 0xab, size); // fault the pages in printf("Footprint after buffer allocation: %5.2f GB\n", footprint_gb()); MTLResidencySetDescriptor * desc = [[MTLResidencySetDescriptor alloc] init]; id<MTLResidencySet> rset = [dev newResidencySetWithDescriptor:desc error:nil]; [desc release]; [rset addAllocation:buf]; [rset commit]; [rset requestResidency]; [rset endResidency]; [rset removeAllAllocations]; [rset commit]; [rset release]; [buf release]; [dev release]; } sleep(5); printf("Footprint 5s after teardown: %5.2f GB\n", footprint_gb()); return 0; } workaround.m: // Build: clang -framework Foundation -framework Metal -o workaround workaround.m #import <Metal/Metal.h> #include <mach/mach.h> // Returns the physical memory footprint of the process. static double footprint_gb(void) { task_vm_info_data_t info; mach_msg_type_number_t n = TASK_VM_INFO_COUNT; task_info(mach_task_self(), TASK_VM_INFO, (task_info_t)&info, &n); return (double)info.phys_footprint / 1e9; } // Performing any work on the GPU ensures the memory from the residency set will be released. static void do_dummy_work(id<MTLDevice> dev, id<MTLCommandQueue> queue) { @autoreleasepool { id<MTLBuffer> tmp = [dev newBufferWithLength:1 options:MTLResourceStorageModeShared]; id<MTLCommandBuffer> cb = [queue commandBuffer]; id<MTLBlitCommandEncoder> enc = [cb blitCommandEncoder]; [enc fillBuffer:tmp range:NSMakeRange(0, 1) value:0]; [enc endEncoding]; [cb commit]; [tmp release]; } } int main(int argc, char ** argv) { printf("Footprint at start: %5.2f GB\n", footprint_gb()); @autoreleasepool { id<MTLDevice> dev = MTLCreateSystemDefaultDevice(); id<MTLCommandQueue> queue = [dev newCommandQueue]; // Workaround that ensures the memory will be released. // It also works if we call this after the residency set release or at any point in between. do_dummy_work(dev, queue); // Allocate ~4GB of memory. const size_t size = 4ULL << 30; id<MTLBuffer> buf = [dev newBufferWithLength:size options:MTLResourceStorageModeShared]; memset(buf.contents, 0xab, size); // fault the pages in printf("Footprint after buffer allocation: %5.2f GB\n", footprint_gb()); MTLResidencySetDescriptor * desc = [[MTLResidencySetDescriptor alloc] init]; id<MTLResidencySet> rset = [dev newResidencySetWithDescriptor:desc error:nil]; [desc release]; [rset addAllocation:buf]; [rset commit]; [rset requestResidency]; [rset endResidency]; [rset removeAllAllocations]; [rset commit]; [rset release]; [buf release]; [queue release]; [dev release]; } sleep(5); printf("Footprint 5s after teardown: %5.2f GB\n", footprint_gb()); return 0; }
Replies
0
Boosts
0
Views
38
Activity
9h