Overhead Prediction for PIM-Enabled Applications with Performance-Aware Behaviour Models

Authors:

Abstract
Today’s processing-in-memory (PIM) devices, such as UPMEM PIM, behave like computational offloading engines: applications must explicitly allocate memory ranks, upload PIM kernels, transfer data and control commands, launch kernels, and read back output data. Past research has focused on identifying offloading candidates and predicting the performance of corresponding PIM kernels. However, this is only half the truth: PIM management and data transfer overheads can easily jeopardize kernel-level performance gains, and thus must be taken into account by scheduling algorithms. These overheads, incurred as part of software development kit (SDK) operations, are application- and workload-dependent: workload partitioning (i.e., the amount and size of data transfers) often depends on workload size and resource allocation. Hence, PIM performance models must account for PIM SDK operations in addition to kernel execution, requiring time-intensive, application-specific benchmarks that must be repeated for each SDK release or workload change. We present a more efficient performance modelling approach based on behaviour models. We train a behaviour model that predicts an application’s PIM SDK operations for arbitrary workloads and resource allocations, and combine it with application-independent performance models for these operations. Thus, each application must be profiled only once, and profiling does not rely on availability of PIM modules or cycle-accurate simulation. We evaluate a proof-of-concept implementation of our approach on twelve UPMEM PIM applications, and show that it can predict application-specific PIM overheads from simulation traces with an error of less than 25 % in nine cases, and no more than 10 % in six cases.
Funding Information
  • Deutsche Forschungsgemeinschaft

This publication has 21 references indexed in Scilit: