On 26 December 2017 at 19:24, kefu chai <tchaikov@xxxxxxxxx> wrote: > xuehan, could you avoid top-posting? otherwise other subscribers will > have trouble to get context of the discussion if he/she fail to read > the whole thread. > > On Tue, Dec 26, 2017 at 2:35 PM, Xuehan Xu <xxhdx1985126@xxxxxxxxx> wrote: >> Thanks for your reply, kefu:-) >> >> I checked /opt/rh/devtoolset-7/root/usr/include/c++/7/x86_64-redhat-linux/bits/c++config.h, >> and saw the following configuration: >> >> # define _GLIBCXX_USE_DUAL_ABI 0 > > yes. this is the crux of the problem, see > https://git.centos.org/blob/rpms!devtoolset-7-gcc.git/2e5ef6c7934d4417e095855478736742b35bd0af/SPECS!gcc.spec#L1026 > . so, > > - “--with-default-libstdcxx-abi=gcc4-compatible”, the DTS-7's > libstdc++'s default ABI is the old one, and > - it does *not* use dual ABI. > > hence what we have is the gcc4 compatible ABI, and nothing more, on centos 7. > >> >> #if ! _GLIBCXX_USE_DUAL_ABI >> // Ignore any pre-defined value of _GLIBCXX_USE_CXX11_ABI >> # undef _GLIBCXX_USE_CXX11_ABI >> #endif >> >> #ifndef _GLIBCXX_USE_CXX11_ABI >> # define _GLIBCXX_USE_CXX11_ABI 0 >> #endif >> >> It seems that no matter how we config MAKEFILE or CMakeFile, >> _GLIBCXX_USE_CXX11_ABI would always be 0 under the circumstance. >> So I think https://github.com/ceph/ceph/pull/18755 is still needed. >> >> Is this right? > > right. unless we swing our own homebrew gcc-7. > >> >> On 26 December 2017 at 00:38, kefu chai <tchaikov@xxxxxxxxx> wrote: >>> On Mon, Dec 25, 2017 at 10:43 PM, kefu chai <tchaikov@xxxxxxxxx> wrote: >>>> On Sat, Dec 23, 2017 at 5:46 PM, Xuehan Xu <xxhdx1985126@xxxxxxxxx> wrote: >>>>> Hi, everyone. >>>>> >>>>> I'm using CentOS 7 with devtoolset-7 to build ceph. However, I found >>>>> that the list::size() in my ceph executables is still counting the >>>>> number of nodes. It's so odd, I googled it, and found that this may >>>>> have something to do with _GLIBCXX_USE_CXX11_ABI. But even if I >>>>> explicitly set _GLIBCXX_USE_CXX11_ABI to 1 when building ceph, the >>>>> problem remains. It seems that _GLIBCXX_USE_CXX11_ABI is set to 0 by >>>>> gcc no matter how I build the source code, even with >>>>> -D_GLIBCXX_USE_CXX11_ABI=1. >>>> >>>> Xuehan, thanks for testing! >>>> >>>> turns out the gcc-7 in devtoolset is compiled with >>>> "--with-default-libstdcxx-abi=gcc4-compatible", see the output of "g++ >>>> -v", so it is using "_GLIBCXX_USE_CXX11_ABI=0", you can by checking >>>> /opt/rh/devtoolset-7/root/usr/include/c++/7/x86_64-redhat-linux/bits/c++config.h >>>> . >>>> >>>> in other words, we still need https://github.com/ceph/ceph/pull/18755. >>> >>> >>> hmm, probably not. >>> >>> add_definitions(-D_GLIBCXX_USE_CXX11_ABI=1) >>> >>> and >>> >>> set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fabi-version=0") >>> >>> might help. I will give it a try tomorrow. > > per the discussion above. we still need it. list::size() is O(n) in > libstdc++ offered by DTS-7. > > > -- > Regards > Kefu Chai Hi, I took a look at https://github.com/ceph/ceph/pull/18755. It seems that it only removes the pg_log.size()'s invocation, however, according to our test, as discusse in https://www.spinics.net/lists/ceph-devel/msg39288.html, there's another place where list::size() can lead to significant performance regression: SimpleLRU, as list::size() is invoked to compute the cache's size when trimming SimpleLRU cache. So I updated the PR https://github.com/ceph/ceph/pull/19397 with a new commit that manually count the size of the cache relying on list::size(). -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
