Performance Regression in QEMU (amd64 Emulating LoongArch64) from 8.0.4 to 9.0.2
Host environment
- Operating system: Arch Linux
- OS/kernel version: Linux lgarch 6.10.3-arch1-1 #1 SMP PREEMPT_DYNAMIC Sun, 04 Aug 2024 05:11:32 +0000 x86_64 GNU/Linux
- Architecture: x86
- QEMU flavor: qemu-loongarch64
- QEMU version: 9.0.2
- QEMU command line:
qemu-loongarch64 $CHROOT/usr/lib/p7zip/7z b
Emulated/Virtualized environment
- Operating system: Arch Linux
- OS/kernel version: N/A
- Architecture: loongarch64
Description of problem
Previous Performance: In May 2023, we were using QEMU 8.0.4 for qemu-user emulation, and the performance was satisfactory. The setup did not include LSX (Loongson SIMD Extensions) support in either the system or QEMU. Performance results are shown in Figure A.
Current Performance: Recently, we upgraded to QEMU 9.0.2. Both the system and QEMU now support vectorized instruction sets. However, we observed a performance decline to less than 60% of the previous benchmarks.
We found that the slowdown is not caused by LSX. Disabling LSX compilation in the new version results in even worse performance. However, there are indeed significant differences between the two systems in terms of LSX support.
Steps to Reproduce:
- Use QEMU 8.0.4 on an x86 machine to run a specific workload
- Upgrade to QEMU 9.0.2
- Run the same workload and compare the performance results
Additional information
We use systemd-nspawn and qemu-binfmt for containerized operations. You can get a clean chroot from lcpu release here
Figure A, performance in May 2023
Figure B, performance nowadays
