verma@ucbarpa.Berkeley.EDU (Dinesh C. Verma) (10/17/89)
I am interested in knowing the impact of fast networks on the architecture and organisation of (futuristic?) computer systems. I would appreciate if you can tell me about any references that deal with this or related topics, or anyone looking into similar areas. Uptil now networks and computer architecture have evolved independently. Part of the reason has been the separation between the I/O architecture and the CPU/MEMORY system. However, since networks are so common nowdays, it is hard to believe that they need not have any impact on the architecture of a machine. I would appreciate any leads. Thanks, Dinesh Verma verma@arpa.berkeley.edu
gmb@iclswe.UUCP (Grahame Budd) (10/18/89)
In article <31963@ucbvax.BERKELEY.EDU> verma@ucbarpa.Berkeley.EDU.UUCP (Dinesh C. Verma) writes: >I am interested in knowing the impact of fast networks on the >architecture and organisation of (futuristic?) computer systems. >...... > Uptil now networks and computer architecture have evolved >independently. Part of the reason has been the separation between >the I/O architecture and the CPU/MEMORY system. However, since >networks are so common nowdays, it is hard to believe that they >need not have any impact on the architecture of a machine. You might like to check out the ICL (U.K.) operating system VME and the hardware that goes with it. The I/O and CPU/MEMORY are totally seperate but the system is linked together *internally* by high speed (250Mbit/sec), local area networks. I do not mean a bus system, this is an intelegent level 4 transport service. In the latest version of the VME Architecture the Disc, intra central suite and 'slow' device systems are all based on LANs. The 'slow' devices run on Ethernet LAN's running at 10 Mbit/sec. The overall system performance is VERY good, the use of networking techniques inside the "CPU" doesn't slow things down and makes totally resilient systems available on a general purpose commercial mainframe. There are no restrictions or special techniques for the software. In most multi systems where the processors are a shared resource to a single operating system there are severe multi-processor overheads because of the problems involved in shared access to i/o controllers and store. Most systems get no better than 1.3 times a single for two processors unless these are segregated and so become two operating systems sharing certain resource. In VME this is not a problem, factors of 1.8 are common. This factor doesn't change much for more processors, i.e. at 4 one can get 3.2 to 3.6 with no problems. Its worth looking out for.