Path: utzoo!mnetor!tmsoft!torsqnt!news-server.csri.toronto.edu!clyde.concordia.ca!thunder.mcrcim.mcgill.edu!snorkelwacker.mit.edu!apple!usc!sdd.hp.com!apollo!hpway!mishkin From: mishkin@jrst.apollo.hp.com (Nathaniel Mishkin) Newsgroups: comp.sys.apollo Subject: Re: NCS Maybe RPC call usage?? Message-ID: Date: 30 Dec 90 23:00:00 GMT References: <9012270444.AA06210@umix.cc.umich.edu> Sender: root@apollo.HP.COM Organization: r_d Lines: 39 In-reply-to: TAYBENGH@NUSDISCS.BITNET's message of 27 Dec 90 20:44 GMT In article <9012270444.AA06210@umix.cc.umich.edu> TAYBENGH@NUSDISCS.BITNET writes: Is Maybe RPC call messages buffered by the system in order to improve throughput, OR it is sent out immediately to achieve low-latency (just as normal SAR RPC call)? The latter. Sounds like what you're interested in is something we've (and others, no doubt) have called "batch mode RPC". The trick in supporting that sort of thing is making it not really ugly to the programmer. You'd really like to just write you interface as usual except that you'd be able to annotate some procedures (in particular ones that have no output parameters) as being "batchable" and other procedures as being synchronization points (i.e., calls to which would causes the system to force out all other batched requests). In any case, we've not implemented this feature. Moreover, does NCS RPC try to optimize intra-machine calls by bypassing expensive& unnecessary data conversion and network communcation operations? Data conversion costs are not paid when machines of like data representation are communicating (assuming they use one of the possible representations as their native representations). Thus, there is no data conversion overhead in intra-machine calls. As to communications costs, we don't do anything special, although we've considered doing so. In NCS 2.0 (part of the forthcoming OSF DCE), which has a somewhat more extensible mechanism with regard to RPC protocol choice, a special path for intra-machine RPC should be easy to add. Of course, it's never so simple to build an intra-machine message passing mechanism (which would be the foundation for RPC) as it seems. By the time you're done, the implementation of the mechanism is moderately hairy and you might have to work hard to make it really faster than the alternatives (e.g., intra-machine IP). -- -- Nat Mishkin Cooperative Object Computing Operation Hewlett-Packard Company mishkin@apollo.hp.com