MITCH LOEBEL COMMENTS ON SCI 10.25.96 FEATURES AND COMMENTARY HPCwire ============================================================================= San Diego, Calif. -- In response to HPCwire's recent three-part interview with Dave Gustavson on SCI (see articles 10249, 10282 and 10316, "DAVE GUSTAVSON ANSWERS QUESTIONS ABOUT SCI, PARTS I, II and III, 10.04.96, 10.11.96 and 10.18.96), B. Mitchell Loebel, strategic alliances and partnering director for the PARALLEL Processing Connection and marketing director for Minute-Tape International Corporation, has written the following note and commentary. --- Whereas Dave chronicled the history and rationale behind SCI, I believe the uninformed reader is not left with any high spots to summarize the issues. As I mentioned, SCI has been plagued with misconceptions (which is why we are improperly compared/contrasted with ATM and DASH). This is a real opportunity to clear away the smoke. HPCwire Question: Briefly describe how SCI operates and contrast this with the other principle networking technologies, such as ATM, HIPPI, etc. LOEBEL: One of the biggest problems that has faced SCI has been the fact that it is not properly understood. Your question itself implies that it is a networking technology, ".... such as ATM, HIPPI, etc." Networking is generally used to move files for non-computational purposes, i.e. sound, graphics, text (such as this note). SCI moves data as part of computational cycles, i.e. loads and stores. Even your question below about moving large packets of information implies non-computational file movement. The DASH system from Stanford University has never gotten stuck with that misconception. In truth, SCI should be contrasted with DASH. Here SCI has several important advantages: 1. SCI is an IEEE standard, whereas DASH is/was a research program, 2. SCI allows for certain optimizations in the cache coherence protocol specifically because the directory of cache copies is based in the caches themselves rather than in memory. I believe that the discussions of SCI's very fast inter-nodal speeds (1 GByte/sec) have skewed public understanding into not seeing it as a COMPUTATIONAL mechanism. Simply put, the SCI protocol enables Distributed Shared Memory architectures. HPCwire Question: What companies presently offer SCI, and how does its cost compare with competing technologies? LOEBEL: The PARALLEL Processing Connection has spawned a start-up called Multi-Node Microsystems Corporation which is developing implementations of the cache coherent interface between both the 604 and P6 buses and a cluster of workstations/PC's tied together with SCI. (The PARALLEL Processing Connection is an entrepreneurial association headquartered in Sunnyvale, founded in late '89. Our raison detre is to be a spawning ground for new high performance computer companies. We hold monthly meetings at Sun Microsystems in Palo Alto and I am its Executive Director.) HPCwire Question: Do you feel SCI receives enough technical-media attention? If not, why? LOEBEL: I am most concerned that the media gives SCI the wrong kind of attention (see above). HPCwire could do the entire computer community a great service by clarifying the issues. HPCwire Question: For what uses is SCI most and least suitable? LOEBEL: SCI has great value where large data sets are involved. In other words, 100s or 1000s of processors needing to access data from a common data set in order to provide many short computations would probably want to access a common shared memory space. Alternatively, each processor would have its own private memory space and the very large data set would have to be moved into each such private memory - that would involve lots of network traffic, and lots of memory for essentially redundant storage. Examples of such applications might include 3D graphics and Virtual Reality rendering, atmospheric modelling as is done at Naval Research Laboratory, and EDA. SCI's other benefit lies in the fact that the shared memory programming model is acknowledged as being far more comfortable to programmers than is explicit message passing. In fact, CC-NUMA which is enabled by SCI is essentially an extension of SMP to very large numbers of processors. HPCwire Question: What can be done to facilitate end-users' understanding of I/O and networking as memory transfers, as SCI requires? LOEBEL: Once again, it is important to understand that SCI involves access to a remote node's memory controller (as in loads and stores) instead of an I/O port. HPCwire Question: Please give a realistic assessment of SCI's future. LOEBEL: SCI offers a sophisticated solution to a COMPUTATIONAL problem (at commodity prices). The problems are the misconceptions noted above and the difficulty in obtaining access to memory/processor buses because most system vendors keep them proprietary. In the case of Intel's P6 bus, it is generally understood that Intel heavy handedly requires a license to be obtained by a vendor of any device which will be plugged into that bus. Worse even is the fact that entrepreneurial enterprises are precluded from participating because such organizations aren't even on Intel's business radar. And all licensees are required to fully disclose their product info to Intel. As you can imagine, this very adversely affects an entrepreneur's opportunity to obtain funding. Parenthetically (and thankfully), it appears that IBM is not taking that route with its 60X bus. As a result, would-be SCI vendors are forced to put their devices on Sun's S-bus or the PCI bus; neither of these buses support cache coherence. Thus, only the high speed interconnect aspect of SCI is widely used and the rest of it is ignored, forgotten, and misunderstood. That will change in the very near future! ---------------- HPCwire welcomes reader comments and suggestions. Please send feedback to Alan Beck, editor in chief, editor@hpcwire.tgc.com ************************************************************************** H P C w i r e S P O N S O R S Product specifications and company information in this section are available to both subscribers and non-subscribers. 936) Sony 905) Maximum Strategy 937) Digital Equipment 934) HP/Convex Tech. Center930) HNSX Supercomputers 932) Portland Group 921) Cray Research Inc. 902) IBM Corp. 915) Genias Software 909) Fujitsu 935) Silicon Graphics **************************************************************************** Copyright 1996 HPCwire. Redistribution of this article is forbidden by law without the expressed written consent of the publisher. For a free trial subscription to HPCwire, send e-mail to trial@hpcwire.tgc.com.