Worker threads allocate at least some recycled cache slots of a local portion of a shared memory to the compute node to which the memory portion is local. More specifically, the recycled cache slots are allocated prior to receipt of the IO that the recycled cache slot will be used to service. The allocated recycled cache slots are added to primary queues of each compute node. If a primary queue is full then the worker thread adds the recycled cache slot, unallocated, to a secondary queue. Cache slots in the secondary queue can be claimed by any compute node associated with the shared memory. Cache slots in the primary queue can be used by the local compute node without sending test and set messages via the fabric that interconnects the compute nodes, thereby improving IO latency.