This patchset is derived from my ongoing work on MTTCG, but does
not depend on it and brings improvements that we can already
benefit from. It applies cleanly on the current master and
is checkpatch-clean.
The key goal is to make the TB hash table faster, and while at it,
scalable. Tested on two different host machines, the execution time
improvement before and after this series, when booting a debian
jessie arm image[*] with arm-softmmu, is:
- Intel Xeon E5-2690: 21.2% less time
- Intel i7-4790K: 23.5% less time
This workload is particularly sensitive to TB hash performance.
Other workloads not as sensitive might see a slight performance
degradation with this patchset, since the hashing + lookup
functions take now more instructions. In any case, no significant
slowdowns should occur.
The commit logs are sometimes long because I have lots of numbers
to share.
The only bits I'm not too comfortable with in this series are patches
2 and 5; I don't develop on Windows so I'm shooting in the dark there.
Please take a look and if possible, test on workloads you care about!
Thanks,
Emilio
[*] taskset -c 0 arm-softmmu/qemu-system-arm -machine type=virt -nographic \
-smp 1 -m 4096 -netdev user,id=unet,hostfwd=tcp::2222-:22 \
-device virtio-net-device,netdev=unet \
-drive file=jessie-arm32.qcow2,id=myblock,index=0,if=none \
-device virtio-blk-device,drive=myblock \
-kernel aarch32-current-linux-kernel-only.img \
-append 'console=ttyAMA0 root=/dev/vda1' \
-name arm,debug-threads=on -smp 1 -tb-size 1024
The image is taken from:
http://people.linaro.org/~alex.bennee/images/jessie-arm32.qcow2
The image was modified to call `shutdown -h now` right after boot.
The kernel is taken from:
http://people.linaro.org/~alex.bennee/images/aarch32-current-linux-kernel-only.img
