Skip to main content
Version: v0.58.0

top_cuda_memory

The top_cuda_memory gadget periodically reports CUDA memory allocation and free activity per process, split by library (libcuda Driver API vs libcudart Runtime API) and by memory type (GPU device memory vs host pinned memory).

It traces alloc/free calls in both libraries independently because the CUDA Runtime (libcudart) internally calls the Driver API (libcuda), and the runtime may cache memory rather than forwarding every free to the driver. Tracking both libraries gives an accurate picture of memory usage from each perspective.

Requirements​

  • Minimum Kernel Version: 5.12 (requires BPF CMPXCHG atomics)
  • CUDA workload using libcuda.so (Driver API) and/or libcudart.so (Runtime API)

Getting started​

Running the gadget:

$ kubectl gadget run ghcr.io/inspektor-gadget/gadget/top_cuda_memory:v0.58.0 [flags]

Guide​

This example shows the gadget running alongside an application that allocates GPU and host-pinned memory using the CUDA Driver and Runtime APIs.

Start the gadget:

$ kubectl gadget run top_cuda_memory:v0.58.0 --pod mypod
K8S.NODE         K8S.NAMESPACE   K8S.PODNAME   K8S.CONTAINERNAME        PID HOST COMM              MEM_ALLOC_BYTES   MEM_FREE_BYTES  MEM_ALLOC_CALLS   MEM_FREE_CALLS  MEM_IMPLICIT_FREE_BYTES  MEM_IMPLICIT_FREE_CALLS

Once a CUDA process starts allocating memory, the gadget will show output like:

K8S.NODE         K8S.NAMESPACE   K8S.PODNAME   K8S.CONTAINERNAME        PID HOST   COMM              MEM_ALLOC_BYTES   MEM_FREE_BYTES  MEM_ALLOC_CALLS   MEM_FREE_CALLS  MEM_IMPLICIT_FREE_BYTES  MEM_IMPLICIT_FREE_CALLS
minikube         default         mypod         mypod                1234567 DEVICE cuda_app               2147483648       1073741824              128               64                        0                        0
minikube         default         mypod         mypod                1234567 HOST   cuda_app                134217728        134217728               16               16                        0                        0

The HOST column distinguishes memory type:

  • DEVICE — GPU device memory (allocated with cuMemAlloc, cudaMalloc, etc.)
  • HOST — Page-locked host memory (allocated with cuMemAllocHost, cudaMallocHost, cudaHostAlloc, etc.)

The gadget reports two separate datasources:

  • libcuda_mem_stats — Driver API (libcuda.so) view
  • libcudart_mem_stats — Runtime API (libcudart.so) view

For applications like ollama that never call cuMemFree explicitly, memory is bulk-freed when the CUDA context is destroyed. In that case MEM_IMPLICIT_FREE_BYTES and MEM_IMPLICIT_FREE_CALLS will reflect the released memory:

K8S.NODE         K8S.NAMESPACE   K8S.PODNAME   K8S.CONTAINERNAME        PID HOST   COMM              MEM_ALLOC_BYTES   MEM_FREE_BYTES  MEM_ALLOC_CALLS   MEM_FREE_CALLS  MEM_IMPLICIT_FREE_BYTES  MEM_IMPLICIT_FREE_CALLS
minikube         default         mypod         mypod                9876543 DEVICE ollama                17179869184                0             1024                0              17179869184                        1

By default the gadget prints a summary each second. Both the number of summaries and the interval can be customized with --map-fetch-count and --map-fetch-interval:

kubectl gadget run top_cuda_memory:v0.58.0 --map-fetch-count 5 --map-fetch-interval 2s