Re: Question about C++ nonblocking server

[David Reiss <dr...@facebook.com>]

> Can anyone shed any light on whether the nonblocking cpp server can do what
> i want? 
I'm sorry, but no it won't.  The TNonblockingServer uses a threadpool
internally, so it still requires one thread per active outstanding request.

--David

Andrei Makhanov wrote:
> Hello,
> 
> I've set up a C++ Server that has several types of jobs that it performs.
> Some of these are very quick lookups that require virtually no processing
> time. Others can take up to several hours of intense cpu processing to
> complete. I am trying to make the service 'nonblocking' in the sense that
> those small processes can always return quickly. At the same time, i want to
> have the ability to run several (predefined number) of these large jobs
> concurrently with new big jobs simply going into a waiting queue. The way my
> system is set up currently, for some calls you dont know how long it will
> take. If the job has been done before, it just returns the precomputed
> results (quickly), but if it had not been done, it does the heavy duty
> processing at that time (can take hours).
> 
> From experimenting with TThreadedServer, it seems it just keeps spinning off
> new threads to handle each request. If there are too many of these long
> jobs, it can potentially bring down the whole server if a bunch of them get
> requested around the same time. This server is always responsive, but very
> prone to abuse.
> 
> The threadpool server makes sure i never kill my machine, however, once it
> has the predefined number of threads running, it stops responding to new
> requests, or rather just sits on them until one of the big jobs completes.
> 
> I've been looking at the Nonblocking server, but with the lack of
> documentation, can't figure out if it will do what i want -- always respond,
> while not killing my machine.
> 
> Can anyone shed any light on whether the nonblocking cpp server can do what
> i want? If so, how do use it? If not, can anyone recommend a different
> solution?
> 
> Thanks,
> andrei