Anthony Liguori a écrit : > On 05/26/2010 09:12 AM, Aurelien Jarno wrote: >>> It's hard for me to consider this a performance regression because >>> ultimately, you're getting greater than bare metal performance (because >>> of extremely aggressive caching). It might be a regression from the >>> previous performance, but that was at the cost of safety. >>> >> For people who don't care about safety it's still a regression. And it >> is a common usage of QEMU. >> > > It's not a functional change. It's a change in performance. There are > tons of changes in performance characteristics of qemu from version to > version. It's not even a massive one. > >>> We might get 100 bug reports about this "regression" but they concern >>> much less than 1 bug report of image corruption because of power >>> failure/host crash. A reputation of being unsafe is very difficult to >>> get rid of and is something that I hear concerns about frequently. >>> >>> I'm not suggesting that the compile option should be disabled by default >>> upstream. But the option should be there for distributions because I >>> hope that any enterprise distro disables it. >>> >>> >> Which basically means those distro don't care about some use cases of >> QEMU, that were for most of them the original uses cases. It's sad. >> > > This isn't a feature. This is a change in performance. No one is not > able to satisfy their use case from this behavior. > >> Sometimes I really whishes that KVM never tried to reintegrate code into >> QEMU, it doesn't bring only good things. >> > > I highly doubt that this is even visible on benchmarks without using > KVM. The improvement on a microbenchmark was relatively small and the > cost from TCG would almost certainly dwarf it.
It is something clearly visible. Before fsync() was not used, and it happens this syscall can be very expensive (ie a few seconds, especially with some other i/o load on the system) on ext3 with not so old kernels. A google search for "firefox fsync" will give you a few pointers. > Also, remember before KVM, we had single threaded IO and posix-aio > (which is still single threaded). If KVM never happened, block > performance would be far, far worse than it is today with cache=writeback. > io thread is not enable by default in QEMU. -- Aurelien Jarno GPG: 1024D/F1BCDB73 aurel...@aurel32.net http://www.aurel32.net
