Performance Impact of Inter-PIM Communication
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionProcessing-in-Memory (PIM) mechanisms explore the application’s data parallelism in different ways. Hence, it is expected that those mechanisms also support thread-level parallelism and data-sharing between threads. However, this can be costly depending on whether the PIM hardware supports data-sharing or relies on the Host. By analyzing data-sharing corner cases, this work shows that this communication, if executed through the Host, can hinder the benefits of PIM devices. We evaluate a PIM-side communication solution that reduces the performance costs of inter-PIM communication by 57%. The solution completely avoids using the host hardware to communicate, significantly improving data-sharing energy efficiency by 11×.