OSDI 2021 papers summary

This post is for recording some notes from a few OSDI'21 papers that I got fun.

DMon: Efficient Detection and Correction of Data Locality Problems Using Selective Profiling

https://www.usenix.org/system/files/osdi21-khan.pdf

DMon is a compiler-based data locality optimization system. The approach is quite similar to that of daphicx or similar things. It injects data access profiling code in the target program, build and run it with some workload, collect the profiling code-generated results, analyze the profile results to figure out what kind of optimization can provide some benefit to the program, injects the optimization code in the program, built it again, and deploy the version to the production.

Biggest novelty of DMon is its selective profiling. It first figure out for what resource (e.g., CPU, L1/L2/L3$, or memory) the program’s performance is bounded, and select what information to collect based on that. It also uses a simple sampling technique. The default sampling rate is 1 sample per 1000 events, though it can be tuned by user. In this way, DMon profile only necessary information and therefore incurs only small overhead (1.36% on average). This allows DMon to do the profiling not only under production-like workloads but also real production.

For the optimization, it uses some existing well-known techniques including (in)direct prefetching, and structure merging, based on the analyzed profiling results. The paper claims this provide 16.83% speedup on average. For the evaluation, the authors use some benchmarks including PARSEC3/SPLASH-2X, NPB, TPC-H on PostgreSQL, and Renaissance benchmark suite.

The idea is novel and makes sense. However, because this is a compiler-based approach, it has a limitation in language. It supports only C and C++ for now. Also, though it claims it can be used on production, I unsure if real production people agrees. Especially, the re-deployment of the optimized version would not be so easy for long-running systems.

Rearchitecting Linux Storage Stack for µs Latency and High Throughput

https://www.usenix.org/conference/osdi21/presentation/hwang

When latency-sensitive apps and throughput-bound apps are co-running, Linux kernel cannot achieve microseconds-scale latency and high throughput. Both SPDK and advanced SPDK usage are same. This paper finds the storage stack of Linux kernel is somewhat like network switch, because it routes I/O request from CPU cores to input queues of storage devices. This paper introduces blk-switch, which modifies the stack to decouple request processing from the application cores and utilizing multiple egress queue having several advanced features including prioritization and steering. It achieves microseconds-scale latency and high throughput under the situation, without modification of the application.

Beyond malloc efficiency to fleet efficiency: a hugepage-aware memory allocator

https://www.usenix.org/conference/osdi21/presentation/hunter

This paper optimizes TCMALLOC by making it more aware of hugepage and using it on Google’s warehouse scale computing. The optimization is mostly based on some model-based heuristics. It’s description on TCMALLOC internal and the problem background is quite useful.

Nap: A Black-Box Approach to NUMA-Aware Persistent Memory Indexes

https://www.usenix.org/conference/osdi21/presentation/wang-qing

Intel released Optane PM DIMM, so real PM world has already started long time ago. This paper explores the performance of the PM indexing on NUMA systems. It finds PM performance is degraded on NUMA systems, and current NUMA-aware PM indexing systems are not working good enough for NUMA systems. Based on findings that most of the real world workloads have highly skewed access pattern, the authors implemented a system called Nap, which translates the PM indexing to be NUMA-awared.