We've measured significant performance improvements for several
benchmarks by using 64K pages (SPECint, sysbench mysql, and kernel
compiling)[*]. I'd therefore like to discuss whether or not we should
switch to 64K pages in vivid.
There's the question of whether or not we would be penalizing the
performance of other classes of workloads people want to run on arm64.
If there are some representative tests we should be looking at, please
let me know.
Also, a known tradeoff is that we'd lose compatibility with existing
ARMv7 binaries. This is because, until very recently, binutils did not
guarantee 64k section alignment for 32-bit binaries:
http://www.spinics.net/lists/arm-kernel/msg383869.html
I'm not sure how much interest there is for ARMv7 compat in Ubuntu.
There is the developer use case of using the same hardware for armhf
and arm64 porting - which, theoretically, we could also achieve by
rebuilding those ports w/ an updated binutils. But otherwise, I'm not
aware of many legacy ARMv7 apps that users are likely to want to bring
over to Ubuntu/arm64 that couldn't be rebuilt.
As points of reference, both OpenSuSE and Fedora appear to be using
64K pages, while Debian is currently using 4K.
[*] Unfortunately I'm not allowed to share these results outside of
Canonical due to NDAs - but I can provide details on the tests we ran
for people who may want to run them on their own hardware.
--
ubuntu-devel mailing list
ubuntu-devel@lists.ubuntu.com
Modify settings or unsubscribe at: https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel
