AMD Threadripper 2950X Offers Great Linux Performance At $900 USD

[Mark Rose]

It's still worth considering. If the threadripper CPU is twice as fast as an intel i7 CPU, would it be worth buying 2 intel-based computers instead? This would be a real-world performance/dollar comparison.

Indeed. If your problem is memory bandwidth bound, then multiple systems with fewer cores is an easy way to get more memory channels.

[GreenReaper]

Did you miss the perf-per-Watt and power numbers? They are towards the end of the article. The minimum power numbers are basically idle.

They aren't there. You need to link to page 10 of the 2990WX benchmarks article, saying there are power per watt graphs for the 2950X there - or stick with a single link to the 2990WX article, but clarify that "more results" means more results for the 2950X across a wider range of apps, in comparison with 2990WX.

As it is, I just thought "this is more results [for another Threadripper]". I didn't link it to being about the CPU being reviewed. Power/watt wasn't mentioned. You actually have great information but people who are focused on the low-end and who know they aren't going to get a 2990WX are likely to miss it.

[pal666]

It's still worth considering. If the threadripper CPU is twice as fast as an intel i7 CPU, would it be worth buying 2 intel-based computers instead?

two computers do not share memory, so not every task could be easily split between two computers
your question would be more correct with multisocket motherboards

[pal666]

Indeed. If your problem is memory bandwidth bound, then multiple systems with fewer cores is an easy way to get more memory channels.

not in the case where you are trading one 4-channel cpu for two 2-channel ones

[Mark Rose]

not in the case where you are trading one 4-channel cpu for two 2-channel ones

You are, of course, correct. I run a lot of mprime (Prime95), which is very memory bandwidth intensive. Two-channel 4-core machines are still usually memory bandwidth bound. Eight cores with only 2 channels or 16 cores with only 4 channels is suboptimal.

[blueweb]

I tried to run the benchmarks on my system (Fedora 28) with

$ phoronix-test-suite benchmark 1808102-RA-AMD2950XT81

but a couple of tests failed to install.

    pts/hpcc-1.2.4:
        Test Installation 1 of 2
        1 File Needed [0.63 MB / 1 Minute]
        File Found: hpcc-1.5.0.tar.gz                               [0.63MB]
        Installation Size: 9 MB
        Installing Test @ 23:16:38

        [NOTICE] Supported install-time optional variables include
        $MPI_PATH, $MPI_INCLUDE, $MPI_CC, $MPI_LIBS, $LA_PATH, $LA_INCLUDE,
        $LA_LIBS, $CFLAGS, $LD_FLAGS, and $MPI_LD

        [NOTICE] Supported run-time optional environment variables include
        $N, $NB, $MPI_NUM_THREADS, $HOSTFILE
            The installer exited with a non-zero exit status.
            ERROR: make[1]: o: Command not found
            LOG: ~/.phoronix-test-suite/installed-tests/pts/hpcc-1.2.4/install-failed.log

    pts/john-the-ripper-1.6.2:
        Test Installation 2 of 2
        1 File Needed [29.36 MB / 1 Minute]
        File Found: john-1.8.0-jumbo-1.tar.gz                      [29.36MB]
        Installation Size: 102 MB
        Installing Test @ 23:16:39
            The installer exited with a non-zero exit status.
            ERROR: JtR requires libssl being installed
            LOG: ~/.phoronix-test-suite/installed-tests/pts/john-the-ripper-1.6.2/install-failed.log


The following tests failed to install:

  - pts/hpcc-1.2.4
  - pts/john-the-ripper-1.6.2

I noticed that the version numbers are different in attempted installation and the failure notice, though not sure if that matters.

Also, libssl does seem to be installed already:

$ locate libssl
~/.phoronix-test-suite/installed-tests/pts/openssl-1.10.0/openssl-1.1.0f/libssl.a
~/.phoronix-test-suite/installed-tests/pts/openssl-1.10.0/openssl-1.1.0f/libssl.pc
~/.phoronix-test-suite/installed-tests/pts/openssl-1.10.0/openssl-1.1.0f/libssl.so
~/.phoronix-test-suite/installed-tests/pts/openssl-1.10.0/openssl-1.1.0f/libssl.so.1.1
~/.phoronix-test-suite/installed-tests/pts/openssl-1.10.0/openssl-1.1.0f/util/libssl.num
/usr/lib/.libssl.so.1.1.0h.hmac
/usr/lib/.libssl.so.1.1.hmac
/usr/lib/libssl.so.1.1
/usr/lib/libssl.so.1.1.0h
/usr/lib/libssl3.so
/usr/lib/rstudio/bin/libssl.so.1.0.0
/usr/lib64/.libssl.so.1.0.2o.hmac
/usr/lib64/.libssl.so.1.1.0h.hmac
/usr/lib64/.libssl.so.1.1.hmac
/usr/lib64/.libssl.so.10.hmac
/usr/lib64/libssl.so
/usr/lib64/libssl.so.1.0.2o
/usr/lib64/libssl.so.1.1
/usr/lib64/libssl.so.1.1.0h
/usr/lib64/libssl.so.10
/usr/lib64/libssl3.so
/usr/lib64/pkgconfig/libssl.pc

Any help?

[GreenReaper]

I tried to run the benchmarks on my system (Fedora 28) with

$ phoronix-test-suite benchmark 1808102-RA-AMD2950XT81

but a couple of tests failed to install.
[...]

[NOTICE] Supported install-time optional variables include
$MPI_PATH, $MPI_INCLUDE, $MPI_CC, $MPI_LIBS, $LA_PATH, $LA_INCLUDE,
$LA_LIBS, $CFLAGS, $LD_FLAGS, and $MPI_LD

[NOTICE] Supported run-time optional environment variables include
$N, $NB, $MPI_NUM_THREADS, $HOSTFILE
The installer exited with a non-zero exit status.
[B]ERROR: make[1]: o: Command not found[/B]

[...]
Any help?

A similar error in something different suggests to me that one of the expected defines (as opposed to the optional ones) has not been defined. You might want to look at the Makefile and see if there are any such defines related to non-installed dependencies. For example, it might expect that CC is defined (containing the name of a C compiler), which it probably won't be if you haven't got a complier installed.

John the Ripper's INSTALL file suggests it uses C (typically the 'gcc' package) and Python as well as packages containing the development headers for libraries such as 'libssl-dev', and possibly C++ (gpp) and things like libnss-dev, libkrb5-dev and libgmp-dev as well (but try without first, I'm not sure what helpers the test actually uses).

[blueweb]

GreenReaper Thanks for the hints.

For the 'hpcc' test, the install-failed.log contains many lines such as:

make[2]: Entering directory '~/.phoronix-test-suite/installed-tests/pts/hpcc-1.2.4/hpcc-1.5.0/hpl/src/auxil/pts'
o HPL_dlacpy.o -c   -I../../../include -I../../../include/pts -I/usr/include -I -fomit-frame-pointer  -O3 -march=native -funroll-loops  ../HPL_dlacpy.c
make[2]: o: Command not found
make[2]: [Makefile:79: HPL_dlacpy.o] Error 127 (ignored)

while the Makefile (in ~/.phoronix-test-suite/installed-tests/pts/hpcc-1.2.4/hpcc-1.5.0/hpl/src/auxil/pts) contains this corresponding line (79) but never defines CC (at least not in that file):

HPL_dlacpy.o           : ../HPL_dlacpy.c           $(INCdep)
    $(CC) -o $@ -c $(CCFLAGS)  ../HPL_dlacpy.c

'gcc' and 'gcc-c++' are installed. Any idea why it's failing? Looks like it thinks 'o' is the command and not the '$(CC)' which should be 'gcc' ?

=====

For 'john-the-ripper', its install-failed.log contains

checking for SSL_library_init in -lssl... no
configure: error: in `~/.phoronix-test-suite/installed-tests/pts/john-the-ripper-1.6.2/john-1.8.0-jumbo-1/src':
configure: error: JtR requires libssl being installed

I checked that the various dependencies are installed:

$ sudo dnf install nss nss-devel krb5-libs krb5-devel gmp gmp-devel
Package nss-3.38.0-1.0.fc28.x86_64 is already installed, skipping.
Package nss-3.38.0-1.0.fc28.i686 is already installed, skipping.
Package nss-devel-3.38.0-1.0.fc28.x86_64 is already installed, skipping.
Package krb5-libs-1.16.1-13.fc28.x86_64 is already installed, skipping.
Package krb5-libs-1.16.1-13.fc28.i686 is already installed, skipping.
Package krb5-devel-1.16.1-13.fc28.x86_64 is already installed, skipping.
Package gmp-1:6.1.2-7.fc28.x86_64 is already installed, skipping.
Package gmp-1:6.1.2-7.fc28.i686 is already installed, skipping.
Package gmp-devel-1:6.1.2-7.fc28.x86_64 is already installed, skipping.
Dependencies resolved.
Nothing to do.
Complete!

There is no 'libssl' in Fedora but there is 'openssl' and 'openssl-devel' which are installed.

[GreenReaper]

There is no 'libssl' in Fedora but there is 'openssl' and 'openssl-devel' which are installed.

May be trying to detect an API removed in openssl-1.1.0. You might want to try uninstalling openssl-dev and installing compat-openssl10-devel with 1.0.2 instead. I'm not a Fedora user but I understand 'dnf install compat-openssl10-devel --allowraising' may do this. Obviously not a great long-term solution.

CC might be defined within the environment by installation of gcc. But maybe it isn't on Fedora and the script-writer didn't know that. You could try defining it yourself, or alternatively modifying the script.

[blueweb]

GreenReaper You're dead on.

Installing the compatibility version 1.0.2 with `sudo dnf install compat-openssl10-devel --allowerasing` allowed john-the-ripper test to install and run successfully. Afterwards, I reinstalled openssl 1.1.0 (`sudo dnf install openssl-devel --allowerasing`) and fortunately the benchmark still works.

For hpcc, I just noticed this notice:

[NOTICE] Supported install-time optional variables include
$MPI_PATH, $MPI_INCLUDE, $MPI_CC, $MPI_LIBS, $LA_PATH, $LA_INCLUDE,
$LA_LIBS, $CFLAGS, $LD_FLAGS, and $MPI_LD

So it's not CC=gcc that needs specifying, but rather MPI-related environment variables.

Installing the test worked with:

env MPI_CC=/usr/lib64/openmpi/bin/mpicc MPI_INCLUDE=/usr/include/openmpi-x86_64/ MPI_LIBS=/usr/lib64/openmpi/lib/libmpi.so MPI_PATH=/usr/lib64/openmpi phoronix-test-suite install-test hpcc

The values for the environment variables were found from the install.sh file over here: https://openbenchmarking.org/innhold...fff99d5828d0ba

The hpcc test is happily crunching away at this very moment.