All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
gcc build time benchmarks on huge-ass dedis
I'm interested in finding out how long it takes to build GCC7 on various big dedis, let's say 6+ cores and 10K+ passmark. It's about 30 minutes on my i7-3770 (4 cores, 9k passmark) but I'd like to see the real monsters, E5-2697, Epyc 32-core and all that. I can supply a test script that runs under Debian and maybe could generate scripts for a few other popular distros, or if you're willing to give me temporary ssh access (non-root, can firewall almost everything) I could run it myself and post the results.
Test duration including setup: under 1 hour
Disk space needed: on order of 1GB
Server should be unloaded during test.
Budget: $0 but I can possibly help you out with some test or dev issue of your own.
Let me know if you want to try this.
Comments
I'd help ya out, we willie, but I have no dedis left, and my 8 core months-to-end-of-contract VPS is burning away at golum precious razorcoins.
I no longer build GCC myself on CentOS as SCL devtoolset yum repos are available. It's how my Centmin Mod LEMP stack can install Nginx and PHP-FPM on CentOS 6/7 with GCC 4.4,4.8, 5.3.1, 6.3.1. 7.2.1 and Clang 3.4 and 4.0.1. Only need to compile for Clang 5.0.1 right now https://community.centminmod.com/threads/testing-gcc-7-1-7-2-clang-4-0-1-scl-yum-packages.12815/
Not that familiar with Debian, but would they have a similar 3rd party repos for such GCC versions as well ?
My OVH MC-64-OC 7700K @ 4.7-5.0Ghz expires in next ~3hrs so can give an Debian OS reload if you can link me to the scripts/instructions to compile and which version of Debian etc
Eva, thanks! Please try the following:
Thanks also to exception0x876, who lent me a 2xE5-2670 16-core VM for the past hour. Build time for the above with -j20 (20 threads, slightly higher than the physical core count) was:
Note that's just over 800% cpu utilization, about half what I'd hope for on a 16 core machine. I saw a lot of the build time was spent re-running configuration scripts or linking .o's. At those times, the make was running essentially single threaded. There was very little time when all 16 cores were in use, and I never saw the 15 minute load average get above 8. The total time beats my i7-3770 but not by a huge amount. Eva2000, your overclocked game machine will probably do better than the E5 because of its higher single thread performance.
I have some other tests that could get better parallelism, if anyone wants to try them sometime (not now).
I'd been wanting to do this test for a while, and it was interesting!
gcc-7 is in buster (testing) and sid on Debian.
But I don't think that's the purpose of this thread
@willie get an error on debian 9.3 64bit
from
GCC
seems this works with
but make fails
Hmm, thanks. Can you try with --disable-multilib? It's odd, libc6-dev-i386 is supposed to supply the 32-bit lib. I wonder if there's some regression in Debian 9.3.
it's compiling with that flag so will see
gcc 7.2.0 built on i7 7700K @ 4.7-5.0Ghz with 2x450GB NVMe raid 1
Thanks!
Yeah gets blocked by config scripts etc a lot of the time.
Not a server, but a machine I have here which is a i7 7820X (8C/16T 3.6-4.5GHz), 32GB RAM, Samsung 960 Pro 512GB, Ubuntu 17.10 amd64 (so GCC 7.2.0),
make -j20
:The compilation software does kinda matter, and really, a proper test needs to use the same version of everything, but anyway...
Build on Dual Xeon E5-2680v4 (28C56T, 2.4G/3.3G Max) with 256GB RAM DDR4 2400MT/s
Build with command
/usr/bin/time make -j64
Thanks xyz and EveNeko. It sounds like clock speed is still king for this test.
E3-1231 (v3 @ 3.40GHz - 3.80GHz)
make -j1
real 131m22.912s
user 121m47.648s
sys 2m9.024s
make -j2
real 69m28.986s
user 124m0.876s
sys 2m17.944s
make -j3
real 49m33.381s
user 128m48.904s
sys 2m26.100s
make -j4
real 40m5.422s
user 133m25.224s
sys 2m27.528s
make -j8
real 35m16.522s
user 207m14.400s
sys 3m35.968s
yes it is. A i7 8700K 6C/12T processor would be ideal