Description
There are currently significant inefficiencies in the handling of memory buffers for large response bodies. When there are many large operations we are seeing significant Large Object Heap (LOH) allocations of byte arrays that are not pooled and reused. Given that the LOH is only garbage collected in Gen2, the cleanup is particularly rare and expensive.
This problem is most prevalent for operation responses > 1MB, though operation responses > 85KB can see some issues on the LOH. The problem is further exacerbated if a Snappy compression algorithm is wired up to the SDK.