Persistent Memory I/O-Aware Task Placement for Mitigating Resource Contention

Hyunwoo Ahn, Jongseok Kim, Euiseong Seo · 2024

Direct access (DAX) file systems for persistent memory (PM) perform reads and writes through load and store instructions, respectively, bypassing the I/O path inside the operating system (OS) kernel. However, because of this, the OS is unable to differentiate PM I/O tasks from CPU-bound tasks, resulting in them being treated equally in task placement. PM I/O-oblivious task placement significantly impacts the DRAM access performance of co-located tasks due to severe resource contention on the memory controller, particularly with remote PM access. Moreover, such task placement fails to utilize on-chip idle resources from the stall cycles for PM I/O, impacting the efficacy of simultaneous multi-threading (SMT). We propose a PM I/O-aware task placement scheme that detects PM I/O activities and dynamically places tasks to mitigate the memory controller contention and to efficiently utilize the idle on-chip resources. In our evaluation, PM I/O-oblivious task placement caused FIO to run over five times slower and reduced SPEC CPU performance by more than three times compared to optimal placements. However, our proposed approach limited the average performance loss to just 3.3% across both workloads, with a maximum loss of only 4.9%.

Read the paper · More papers on PaperTik