[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Increased execution time with TCI in latest git master (was: Re: [PU
From: |
Stefan Weil |
Subject: |
Re: Increased execution time with TCI in latest git master (was: Re: [PULL 00/46] tcg patch queue) |
Date: |
Sun, 7 Feb 2021 23:00:57 +0100 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101 Thunderbird/78.7.0 |
On 07.02.21 19:37, Richard Henderson wrote:
> On 2/7/21 2:50 AM, Stefan Weil wrote:
>> Your latest code from the rth7680/tci-next branch is twice as fast as my code
>> with BIOS boot and qemu-x86_64 on sparc64. That's great.
>>
>> With that code I don't get any BIOS output at all when running qemu-i386.
>> That's not so good.
>>
>> Did I test the correct branch? If yes, I could try the same test on amd64 and
>> arm64 hosts.
>
> Yes, tci-next is the correct branch. I've just rebased it against master,
> which includes the first 30-odd patches.
>
> What host do you not see bios output from qemu-system-i386 (I assume that's a
> typo above)? I see correct output on x86_64, sparc64, ppc64le, and aarch64
> hosts.
Right, the TCI test was done with qemu-system-i386 of course.
I repeated the TCI test with qemu-system-i386 and qemu-system-x86_64 and
the rebased branch.
The system emulation for a BIOS boot works on Apple M1 arm64 with less
that 5 s user time (similar fast as before the latest TCI changes):
./qemu-system-i386 --nographic
4,28s user 0,03s system 37% cpu 11,398 total
./qemu-system-x86_64 --nographic
4,39s user 0,03s system 34% cpu 12,982 total
The same test shows similar timings on an AMD64 server:
./qemu-system-i386 --nographic
user 0m4,958s before tcg-next, 0m5,115s after tcg-next
./qemu-system-x86_64 --nographic
user 0m4,967s before tcg-next, 0m5,263s after tcg-next
Here tci-next is slightly slower than the old code.
The results on sparc64 did not change with the rebased tci-next:
qemu-system-i386 still fails to run, and qemu-system-x86_64 takes about
20 s user time.
Stefan
- [PULL 44/46] accel: extend AccelState and AccelClass to user-mode, (continued)
- [PULL 44/46] accel: extend AccelState and AccelClass to user-mode, Richard Henderson, 2021/02/05
- [PULL 43/46] cpu: tcg_ops: move to tcg-cpu-ops.h, keep a pointer in CPUClass, Richard Henderson, 2021/02/05
- [PULL 45/46] accel: replace struct CpusAccel with AccelOpsClass, Richard Henderson, 2021/02/05
- [PULL 46/46] accel: introduce AccelCPUClass extending CPUClass, Richard Henderson, 2021/02/05
- Re: [PULL 00/46] tcg patch queue, Peter Maydell, 2021/02/06
- Increased execution time with TCI in latest git master (was: Re: [PULL 00/46] tcg patch queue), Stefan Weil, 2021/02/06