Current implementation adopts parallelism at the level of batches. This is fine, however, an even more efficient utilization of resources would be one where we employ Hogwild-like scheme, where idle threads start processing subsequent batches, included in a given tmp aggregation (overwriting happens here). This has potential for much faster ranking of higher-order interactions.
Current implementation adopts parallelism at the level of batches. This is fine, however, an even more efficient utilization of resources would be one where we employ Hogwild-like scheme, where idle threads start processing subsequent batches, included in a given tmp aggregation (overwriting happens here). This has potential for much faster ranking of higher-order interactions.