Just for fun, because I know this is a very contrived test! I wrote a C
program to read/writes block to an zFS file on a z114 z/OS system
connected via FICON to an HDS SAN and Ubuntu server on a Dell PowerEdge
blade server
writing to SAS disks on the rack. Of course, there are latency
differences and my program is probably quite lame. The z/OS was idle at
the time as was the Linux system. I would imagine if they were both
running at full capacity
with high throughput sustained I/O the results would be very different.
#include <stdio.h>
#include <stdlib.h>
int main()
{
size_t numBlocks = 100000;
char * filename = "./io.temp";
#define BLOCK_SIZE 1024
FILE * fp = fopen( filename, "w+" );
if ( fp == NULL )
{
perror( "fopen" );
exit( 8 );
}
char buffer[BLOCK_SIZE] = {0};
for ( int j = 0; j < numBlocks; j++ )
{
if ( fwrite( buffer, sizeof buffer, 1, fp ) == 0 )
{
perror( "fwrite" );
exit( 8 );
}
}
rewind( fp );
while ( fread( buffer, 1, sizeof buffer, fp ) )
{
continue;
}
remove( filename );
return 0;
}
z/OS:
DOC:/u/doc/src: >time iospeed
real 0m 1.15s
user 0m 0.62s
sys 0m 0.20s
Dell:
davcra01@cervidae:~$ time ./iospeed
real 0m0.254s
user 0m0.048s
sys 0m0.199s
On 21/03/2016 10:43 PM, Steve Thompson wrote:
A few years ago, IBM took a Power system and a z/Architecture system
and configured them as closely as they could.
As I recall, they both had the same amount of C-Store available to the
operating system, and they both had the same number of channels (8 if
I remember correctly), and they ran to equivalently sized RAID boxes.
And, if I remember correctly, they were using programs written in C,
that were ported from the one to the other.
The object was to check out how efficiently I/O was prosecuted (done).
The z/Architecture machine finished long before the power system.
I thought I had a copy of that comparison, but I just can't find it so
I can give a link to it.
Regards,
Steve Thompson
On 03/21/2016 09:40 AM, David Crayford wrote:
On 21/03/2016 9:14 PM, R.S. wrote:
Well,
I observed 1,3M IOPS on EC12 or z196 machine during WAS
installation. With minimal CPU utilisation (I mean regular CPU,
I haven't checked SAP).
IMNSHO a PC server with collection of new shining Emulex cards
has waaaay worse I/O capabilities.
We did some tests of database operations on PC. Effects are
unequivocal.
I admire you honesty :) What class of PC server? Was it connected
to a SAN and did it offload I/O to a peripheral device?
BTW: Typical z/OS I/O workload is very different from PC
workload. Much less IOPS, much more data, much less CPU%.
My wife used to work for HDS and I had some interesting
conversations with some of the engineers she used to work with.
That was a few years ago but in their opinion the
high end *nix servers could match a mainframe for I/O throughput.
PC commodity hardware is different but racked up with enterprise
kit I would be interested to know how
they would shape in a drag race. How would a Dell blade with
Infiniband, SAN and enterprise class HBAs compare?
----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO
IBM-MAIN
----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN
----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN
