|
Sorry for how long
it took to reply to your mail. Hope it wouldn't appear to have been left
unnoticed! :-)
HI,
I have done a few IOCP Winsock
application, which work fine, but want to as a couple of things:
I have only designed using for TCP/IP protocol,
and for network, and internet access. I learned from the 'Network Programming
Form Microsoft Windows" book (second edition), and it explains the IOCP really
well, I have in the past read with interest some different comments, and
wondering what people thought of these?
The NumberOfConcurrentThread within the
CreateIOCompletionPort(..) indicates the number of threads allowed to execute
concurrently to avoid thread context switching, which seems logical to me. So
setting it to 0, will inform the processor to allow as many threads to run as
processors.
Well, the rule of thumb here is to have 2 times the number of CPUs
on your machine IOCP worker threads. With
IOCP, the system issues IOCP services to the thread which had been doing it most
recently. Thus, if your thread is awakened to service an IOCP request, and
it finishes up before its time slice expires, and there's another request
available to service, it gets a chance to service them. This means less
context switch, and more performance.
Now the worker threads depends on the type
of application. In my next application there will be database acess from users
connected to the server. So, It also mentions it is possible to have 'x'
amount of worker threads created, but only one will be accessed on the
Completion port at one time. But the handy part is that , which in my case,
if database is being accessed, then there is a chance that the worker
thread be suspended doing its thing with the database, and passing information
to the client application. In this case, another worker thread will be allowed
to operate in its place, which also sounds fair enough. So in my application,
because of using databases, I am thinking of creating say, 5-6 worker threads.
Even though only one will accessing the IOCP routine, another one can take
over if it becomes busy with database access. I am only planning to post one
WSARecv(..) per socket handle, so not much chance of receiving blocks at
different times.
You need to notice a subtle point. No matter
how many threads you have, the system can execute only one thread per CPU,
so in fact the impression that "another one can take over if it becomes busy"
is not quite correct. Maybe letting it run and service other
requests as well can lead to better
performance.
Also, I understand that Per I/O operation
structure, as it is only used for sending and receiving, but, when a
connection is established from a client, a WSARecv(..) is posted, which
has to have a pointer to a Per I/O structure. And within the IOCP routine, the
pointer to this buffer is returned with the data in it, then another
WSARecv(..) is posted within the worker thread. As I can see it, then while
the socket connection is live, each connectoin will alway have a Per I/O
attatched to it. So if I have 500 connections, each one will have its own Per
I/O structure , as it is passed through the WSARecv(...). Is this
correct??
Yes. As their name says, per I/O structures
exists one per each outstanding I/O operation. A big performance
boost is gained if you don't allocate each per I/O structure when
you want to issue an I/O. It would be mich better if you preallocate a big
pool of per I/O objects, and just grab one from the pool when you want
to issue an I/O, and put it back on the pool when an I/O finishes.
|
