samskalicky commented on a change in pull request #20043:
URL: https://github.com/apache/incubator-mxnet/pull/20043#discussion_r613507427
##########
File path: src/storage/pooled_storage_manager.h
##########
@@ -148,9 +157,21 @@ void GPUPooledStorageManager::Alloc(Storage::Handle*
handle) {
mxnet::common::cuda::DeviceStore device_store(handle->ctx.real_dev_id(),
true);
size_t free, total;
cudaMemGetInfo(&free, &total);
- if (free <= total * reserve_ / 100 || size > free - total * reserve_ / 100)
+ double mem_limit_in_bytes = total * memory_limit_percentage_ / 100.0;
+ free = mem_limit_in_bytes - used_memory_;
+ if (free <= mem_limit_in_bytes * reserve_ / 100 ||
+ size > free - mem_limit_in_bytes * reserve_ / 100)
ReleaseAll();
+ if (used_memory_ + size > mem_limit_in_bytes) {
+ // This calls abort() unless
+ // DMLC_LOG_FATAL_THROW != 0, then it
+ // throws std::runtime_error()
+ LOG(FATAL) << "memory limit reached, used: " << used_memory_ << "
limit: "
+ << mem_limit_in_bytes << " (" << memory_limit_percentage_ <<
"% of "
+ << total << ")";
+ }
Review comment:
This is the enforcement of the user-specified GPU memory limit. If users
say "MXNet can use 1GB", then it needs to someone let users know that it cant
function properly anymore within the set memory limit.
Its not any different than running on a GPU with small amount of GPU memory
than needed (artificial limit).
--
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.
For queries about this service, please contact Infrastructure at:
[email protected]
