Bad performance multithread

[jmartin]

Hi,

I have made a multithread version of a program (load a file into
database), and with two processors I get the double of time in the
multithread than in the process (unithread) version.

I have done the test of creating a unique thread that does the same
code that the process version of the program and it takes more time
that the process without creating the thread.

It seems that the thread get worse performance that the process alone,
and that two processors isn't good enough to get a performance
improvement with multithread overhead.

Is this normal? The program requires threads syncronization and that
can get performance worse.

We are using RogueWave threads module under AIX.

Thanks in advance,

[P.J. Plauger]

"jmartin" <jo***************@cgi.comwrote in message
news:11**********************@p79g2000cwp.googlegr oups.com...

I have made a multithread version of a program (load a file into
database), and with two processors I get the double of time in the
multithread than in the process (unithread) version.

I have done the test of creating a unique thread that does the same
code that the process version of the program and it takes more time
that the process without creating the thread.

It seems that the thread get worse performance that the process alone,
and that two processors isn't good enough to get a performance
improvement with multithread overhead.

Is this normal? The program requires threads syncronization and that
can get performance worse.

We are using RogueWave threads module under AIX.

You don't introduce multithreading to improve *performance* but to
improve *responsiveness*. There's a difference. Having said that, I
observe that performance losses as bad as you describe can usually
be dramatically improved by a more judicious use of synchronizing
logic. (Watch where you put your thread locks.)

P.J. Plauger
Dinkumware, Ltd.
http://www.dinkumware.com

[Roberto Waltman]

"P.J. Plauger" <pj*@dinkumware.comwrote:

>"jmartin" wrote in message

>I have made a multithread version of a program ...

and with two processors I get the double of time in the

>multithread than in the process (unithread) version.
...

You don't introduce multithreading to improve *performance* but to
improve *responsiveness*...

That's correct in uniprocessor systems. The OP mentions "two
processors", in which case some programs (not all) can get a
performance boost by multithreading. (But of course, you know that.)

To the OP: try also posting in comp.programming.threads

[Tom St Denis]

jmartin wrote:

Is this normal? The program requires threads syncronization and that
can get performance worse.

Are you asking if two parallel threads competing for the same resource
will be less efficient then one thread using the resource fully?

YES IT WILL BE SLOWER

Tom

[websnarf]

jmartin wrote:

I have made a multithread version of a program (load a file into
database), and with two processors I get the double of time in the
multithread than in the process (unithread) version.

I have done the test of creating a unique thread that does the same
code that the process version of the program and it takes more time
that the process without creating the thread. [... etc]

Optimizing algorithms on multithreaded systems can be hard. If you
split an operation on an array, for example, then you will need each
processor to have half of the memory from that array, which means that
inevitably which may have started as a single continguous array has to
be at least half copied into the cache of the other processor. This is
ok if your algorithm is very expensive in comparison to the cost of
copying the array, for example. But its something trivial like just
summing up the elements of the array, I could see how it could easily
be slower, or at least no faster, on a single threaded machine.

Optimization on multithreading system, still relies good algorithm
design and the use of tools such as profilers to isolate performance
bugs. In that sense, its the same as single threading. But you have
to be aware of hidden performance problems such as increased memory
locality issues such as I have described above.

--
Paul Hsieh
http://www.pobox.com/~qed/
http://bstring.sf.net/

[P.J. Plauger]

"Roberto Waltman" <us****@rwaltman.netwrote in message
news:fs********************************@4ax.com...

"P.J. Plauger" <pj*@dinkumware.comwrote:

>>"jmartin" wrote in message

>>I have made a multithread version of a program ...

and with two processors I get the double of time in the

>>multithread than in the process (unithread) version.
...

You don't introduce multithreading to improve *performance* but to
improve *responsiveness*...

That's correct in uniprocessor systems. The OP mentions "two
processors", in which case some programs (not all) can get a
performance boost by multithreading. (But of course, you know that.)

Yes, I almost added that qualifier. At the risk of quibbling, however,
I still feel obliged to point out that one program can get better
*response* (quicker termination) by using two or more processors, at
the expense of using somewhat more total processing time (worse overall
*performance*). On a heavily loaded system, you're simply stealing from
the commonweal. On a system dedicated to one person, and under utilized,
you can arguably define "performance" in a more selfish manner.

P.J. Plauger
Dinkumware, Ltd.
http://www.dinkumware.com

[Kevin Sproule]

"jmartin" <jo***************@cgi.comwrote in message
news:11**********************@p79g2000cwp.googlegr oups.com...

Hi,

I have made a multithread version of a program (load a file into
database), and with two processors I get the double of time in the
multithread than in the process (unithread) version.

I have done the test of creating a unique thread that does the same
code that the process version of the program and it takes more time
that the process without creating the thread.

It seems that the thread get worse performance that the process alone,
and that two processors isn't good enough to get a performance
improvement with multithread overhead.

Is this normal? The program requires threads syncronization and that
can get performance worse.

We are using RogueWave threads module under AIX.

Thanks in advance,

Have you determined that the process (single or multi-threaded) is CPU bound
and not I/O bound? If it is I/O bound then running multiple threads would
make it more I/O bound and therefore slower.

Kevin Sproule