Re: optimizing link between client and server

["Amitabh A. Nene" <anene cup hp com>]

Yeah, that was the kind of information I was looking for. Things like how
much more expensive would a call within the same address space be compared
to a virtual call. Is it exactly as much as a virtual call or does it do
additional checks - will appreciate pointers to source files that
implement call linking - at this point I am most concerned about the same
address case since I'm thinking of using CORBA just for its interfaces so
that I'll be scale my components to use different transports in the
future.

Thanks
-Amit


On Thu, 30 Nov 2000, Simon Britnell wrote:

> Amitabh A. Nene wrote:
> 
> > I wanted to know how does one go about optimizing the link between a CORBA
> > client and server using ORBit / any other ORB in general. In other words,
> > how does one configure the link depending on whether the client and server
> > are in the same address space, different address spaces on the same
> > physical machine, or on separate physical machines ?
> 
> Hmmm.  I'm a little unsure of what you're asking.  Your question appears 
> to be "How do I select which transport an ORB is using (SHM,TCP/IP,Unix 
> Sockets,Sun RPC,etc) for any ORB?" The answer is "That depends on the 
> ORB." ORBit specifically has a configuration file /etc/orbitrc which 
> switches between unix sockets and tcp/ip as these are the transports it 
> provides.".  Many ORBs automatically make the decision to use 
> dync-call/no marshalling whenever they detect that they are both the 
> client and the server for an IOR.  I *think* ORBit is one of these.
> 
> It also occurs to me that your real issue may be a performance one, in 
> which case I can only say "Minimise round trips".  The further the 
> distance between your client and server, the bigger the latency penalty 
> you will pay.  Calling a remote object method fetchNextFoo() repeatedly 
> to fetch a thousand records will perform fine over SHM where latency is 
> <1ms, but will die screaming over the internet even at a latency of only 
> 100ms (giving a 10 second response time for 1000 records in latency 
> alone) and >500ms latencies are not uncommon.  Far better to implement a 
> method fetchAllFoo() which is called once.
> 
> > Also, in the current ORBit implementation, do these paths translate to
> > dyn-call/no-marshaling, shared-mem, and RPC respectively, if at all ?
> 
> I don't think ORBit does SHM yet.  I know it does dyn-call, unix 
> sockets, tcp/ip4 and tcp/ip6 (same ram address space, same machine, 
> seperated by network).
> 
> Hope this helps.
> 
> 
> _______________________________________________
> orbit-list mailing list
> orbit-list gnome org
> http://mail.gnome.org/mailman/listinfo/orbit-list
>