| Full text | |
| Author(s): |
do Nascimento, Rodrigo Pereira
;
Vel Lejbman, Alfredo Goldman
Total Authors: 2
|
| Document type: | Journal article |
| Source: | 2024 INTERNATIONAL SYMPOSIUM ON COMPUTER ARCHITECTURE AND HIGH PERFORMANCE COMPUTING WORKSHOPS, SBAC-PADW; v. N/A, p. 8-pg., 2024-01-01. |
| Abstract | |
Tracing file I/O operations in Linux is critical for understanding and optimizing application performance, particularly in I/O-intensive environments. However, the choice of tracing tool can dramatically impact the application throughput. This paper investigates the performance overhead associated with two widely used tracing tools - strace and bpftrace - when applied to file I/O operations. Through empirical evaluation using a controlled benchmark over 100 epochs, we compare the throughput distributions of training a machine learning model under both tools against a baseline without tracing. The results reveal that strace introduces significant performance degradation, reducing throughput by 17%. In contrast, bpftrace, leveraging the efficiency of BPF, reduces the throughput by only 1% compared to the baseline, while maintaining consistent performance across trials. (AU) | |
| FAPESP's process: | 19/26702-8 - Trends on high performance computing, from resource management to new computer architectures |
| Grantee: | Alfredo Goldman vel Lejbman |
| Support Opportunities: | Research Projects - Thematic Grants |