On Thu, 4 Sep 2014 09:54:39 +0000
PROT Pierre-François <[email protected]> wrote:
> Hello list,
>
> We have currently migrating part of our production appl from
> Unix/Pseries to zLinux/zEC12. We have ported microfocus cobol
> programs (previously running on PSeries) on our zEC12/zLinux
> plateform and after comparing CPU times on z and on P, it turns out
> that overall cpu times (java and oracle) are much better on the Z but
> it also turns out that the cpu times for OpenCobol are far worse
> (although time elapse treatment remain best on Z).
>
> As OpenCobol generates a C program from the Cobol source code we
> investigated on th C side. After some days of analysis it seems that
> on zLinux, the C calls to DLLs are more expensive on z than on x86 or
> PSeries architecture. The application of different optimization
> options (-O3, march, mtune) does not change the results. Finally, a C
> program that does not calls to external modules is much more
> efficient on zLinux. Here is our test program with libgmp :
>
> #include <stdio.h>
> #include <stdlib.h>
> #include <gmp.h>
>
> int main(void)
> {
> mpz_t x;
> mpz_t y;
> mpz_t result;
>
> long int i=0;
>
> while (i++<20000000)
> {
>
> mpz_init(x);
> mpz_init(y);
> mpz_init(result);
>
> mpz_set_str(x, "7612058254738945", 10);
> mpz_set_str(y, "9263591128439081", 10);
>
> mpz_mul(result, x, y);
>
> mpz_clear(x);
> mpz_clear(y);
> mpz_clear(result);
> }
>
> return EXIT_SUCCESS;
> }
>
>
> It was compiled with gcc testgmp.c -lgmp
> and runned by the command time ./a.out to obtain CPU times. the
> results are the following :
> zLinux ZEC12/5.5 gHz
> real 0m8.930s
> user 0m8.802s
> sys 0m0.006s
>
> linux on x86 HP Blade/2.2gHz
> real 0m7.464s
> user 0m7.438s
> sys 0m0.000s
>
>
>
> When we make trace on library calls, zlinx seems better :
>
> zlinux
> ltrace -c ./testgmp
> % time seconds usecs/call calls function
> ------ ----------- ----------- --------- --------------------
> 41.05 95.646614 159 600000 __gmpz_init
> 29.01 67.588020 112 600000 __gmpz_clear
> 20.27 47.224456 118 400000 __gmpz_set_str
> 9.68 22.561771 112 200000 __gmpz_mul
> 0.00 0.000058 58 1
> ------ ----------- ----------- --------- --------------------
> 100.00 233.020919 1800001 total
>
>
> linux on x86
> ltrace -c ./testgmp
> % time seconds usecs/call calls function
> ------ ----------- ----------- --------- --------------------
> 34.49 153.162784 255 600000 __gmpz_init
> 32.22 143.106564 238 600000 __gmpz_clear
> 25.08 111.395971 278 400000 __gmpz_set_str
> 8.21 36.475142 182 200000 __gmpz_mul
> 0.00 0.000119 119 1
> ------ ----------- ----------- --------- --------------------
> 100.00 444.140580 1800001 total
>
>
>
>
> We also conducted tests on Cobol on different zlinux distributions
> but the problem seems to be the same for all z platforms. (Results
> below: 6m40 for zLinux and 4m23 for x86). On our zLinux, the system
> disk / is on CKD devices, so we also conducted a test with the / on
> FCP disk to see if the problem could come from CKD disks. The results
> is below: there is an improvement in CPU time (6m15 )but still worst
> than on x86. RHEL 6.5 on zEC12 (with CKD devices for system) real
> 7m5.370s user 6m41.487s sys 0m1.166s RHEL 6.5 on zEC12 with
> FCP devices for the system real 7m47.700s
> user 6m14.521s
> sys 0m1.061s
>
> CentOS 6.4 on zEC12
> real 11m35.152s
> user 6m40.513s
> sys 0m1.094s
>
> Debian 7.4 on zEC12
> real 16m1.261s
> user 6m38.751s
> sys 0m1.136s
>
> CentOS 6.4 on x86
> [zadmin@QAVTL318041 ~]$ ps -aef | grep TST
> zadmin 19327 19317 90 14:54 pts/0 00:04:22 /TSTPAT/bin/cobcrun
> TSTSEQREW real 5m25.090s
> user 4m21.355s
> sys 0m1.647s
>
>
> We also conducted tests with a recompiled gcc 4.9.1 but results are
> the same.
>
> We finally used the benchmark tool provided by libgmp on our 3
> platerforms and the results are the following :
>
> RHEL65 zlinux with gmp 4.3.1 : 550.71
> RHEL65 x86 with gmp 4.3.1 : 1960.4
> Debian 7.4 zLinux with gmp 5.0.5 : 1291.9
>
> Bench results is 2 times better on debian than on REHL (the hardware
> is the same and the zVM guest are the same too) and HP blade 2.2GHz
> is also better than zEC125.5GHz :-( It's seems that there is a
> problem on RHEL65 s390x with libgmp (and maybe others libs). Does
> anyone has always seen this problem or as any ideas on how to solve
> this problem ?
looking into the gmp sources it seems that 4.3.1 doesn't have a proper
support for s390x, so it's not optimized for 64-bit s390 at all, while
5.0.5 already has some. Version 5.1.1 from RHEL-7 is better and version
6.0.0 is even much better, they already contain optimized routines
written in assembler. I think this also explains the performance
compared to x86/x86_64 as they have optimized routines already in 4.3.1
(and earlier).
Please open a request for rebase of gmp via your Red Hat support
channel. I don't know whether the rebase is feasible, but we should at
least track the problem.
Dan
----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390
----------------------------------------------------------------------
For more information on Linux on System z, visit
http://wiki.linuxvm.org/
