Development for dual core machine

[Andy]

Hi guys,

I'm sorry, I'm not sure this is the correct group to be asking this
kind of question...

I'm going to develop a software package that includes a web server
(and PHP code) , a database, and some Python code of course. I am
being strongly suggested to make it to work on a dual- or multi-core
computer, but I'm confused on how to take advantage of the multiple
CPUs.

From what I read, I think that simply by making the package run in

several separate processes (web server, database server, Python
interpreter, etc.), and/or using multiple threads (which I will
anyway) the package should be able to use multiple CPUs.

Is my understanding correct? So actually, all I have to do is just
write my multi-threaded Python code as usual? And how is it decided
which process and which threads go to CPU 1, CPU 2, etc.? Perhaps the
BIOS?

Any advice greatly appreciated.
Andy

 

[Bjoern Schliessmann]

I'm going to develop a software package that includes a web server
(and PHP code) , a database, and some Python code of course. I am
being strongly suggested to make it to work on a dual- or
multi-core computer,

No problem. CPython will work on any dual core CPU.

but I'm confused on how to take advantage of the multiple CPUs.

First: Use a web server that can make use of multiple cores.

Second: Use a data base that can make use of multiple cores.

Third, for using multiple cores from CPython: this is an FAQ, please
look at the mailing list archives. Whether you should optimize your
python application to use all cores strongly depends on what your
Python application actually does.

From what I read, I think that simply by making the package run in
several separate processes (web server, database server, Python
interpreter, etc.), and/or using multiple threads (which I will
anyway) the package should be able to use multiple CPUs.

CPython has the GIL (global interpreter lock). Please search for it
in the archives, it's been discussed exhaustingly.

And how is it decided which process and which threads go to CPU 1,
CPU 2, etc.? Perhaps the BIOS?

No, the kernel (i. e., Linux). The BIOS is completely out of this.

Regards,


Björn

 

[samwyse]

Hi guys,

I'm sorry, I'm not sure this is the correct group to be asking this
kind of question...

I'm going to develop a software package that includes a web server
(and PHP code) , a database, and some Python code of course. I am
being strongly suggested to make it to work on a dual- or multi-core
computer, but I'm confused on how to take advantage of the multiple
CPUs.

several separate processes (web server, database server, Python
interpreter, etc.), and/or using multiple threads (which I will
anyway) the package should be able to use multiple CPUs.

Is my understanding correct? So actually, all I have to do is just
write my multi-threaded Python code as usual? And how is it decided
which process and which threads go to CPU 1, CPU 2, etc.? Perhaps the
BIOS?

Any advice greatly appreciated.
Andy

The Python interpreter is not multi-cpu aware, so using Python threads
won't work on multiple CPUs. If your tasks are CPU-bound, then fork
multiple processes. Most web servers (Apache) can handle this
automatically for you.

 

[Bryan Olson]

I'm going to develop a software package that includes a web server
(and PHP code) , a database, and some Python code of course. I am
being strongly suggested to make it to work on a dual- or multi-core
computer, but I'm confused on how to take advantage of the multiple
CPUs.

From what I read, I think that simply by making the package run in
several separate processes (web server, database server, Python
interpreter, etc.), and/or using multiple threads (which I will
anyway) the package should be able to use multiple CPUs.

Right. There is a theoretical possibility that Python's GIL
could limit thread parallelism, but in this kind of application
it should not matter in the least. Either you're I/O bound, or
there are plenty of tasks the operating system could schedule.

If you have the option to run the Python interpreter within a
web-server process, that's usually a performance winner.

Is my understanding correct? So actually, all I have to do is just
write my multi-threaded Python code as usual? And how is it decided
which process and which threads go to CPU 1, CPU 2, etc.? Perhaps the
BIOS?

The O.S. actually. A lot of really smart people have put a whole
lot of work into making the O.S. do that well.

If you usually write your Python apps multi-threaded, as I do,
that's fine. Multi-core efficiency has only a little to do with
it. Web service are, for the most part, intrinsically
parallelizable: run multiple web servers, and multiple Python
interpreters serving multiple connections.

The only hard part is shared data. Scalability is all about
the database.

 

[Paul Rubin]

several separate processes (web server, database server, Python
interpreter, etc.), and/or using multiple threads (which I will
anyway) the package should be able to use multiple CPUs.

Python threading doesn't support multiple CPU's because of the GIL.

One approach to using your multiple cores is to embed a Python
interpreter into a web server that runs in multiple processes,
e.g. mod_python under Apache.

Another is write your application as a separate process (e.g. as an
FastCGI), then run multiple instances of it, connected to a concurrent
web server. For what I'm currently doing, we're using lighthttpd as
the http server and flup to connect to a set of Python FCGI's.

For that matter, if your machine is just dual core, maybe it's ok to
just run a single Python process, figuring that will run on one core
and your database/httpd will run on the other one. However, you
should figure the dual core situation won't last. Dual socket
motherboards are fairly cheap these days, so we have a number of
4-core machines (two AMD dual core cpu's); Intel is already shipping
quad core cpu's, so that puts 8 cores in a dual socket board (you can
already buy Macintoshes configured that way). Higher end server
motherboards have 4 sockets, so you have to expect that 16-core
servers will be common pretty soon.

So if you're working on a cpu-intensive application it's worth your
while figuring out how to parallelize it.

Note that the most careful concurrency stuff probably is in the
database. Serious ones already know how to use multiple CPU's.

 

[Bjoern Schliessmann]

The Python interpreter is not multi-cpu aware, so using Python
threads won't work on multiple CPUs.

Are you sure about this?

Regards,


Björn

 

[Bjoern Schliessmann]

Python threading doesn't support multiple CPU's because of the
GIL.

:s/support/take full advantage of/

Regards,


Björn

 

[Andy]

Thanks guys for the suggestions.

Andy

 

[Willi Richert]

Thanks guys for the suggestions.

Andy

It might be that you have to set the CPU affinity for your python process so
that it works at all. I had that problem on a dual core machine with
hyperthread enabled. Using taskset
(http://www.linuxcommand.org/man_pages/taskset1.html) helped solving the
problem:

taskset -c <CPU#> python ...

Best regards,
wr