With processors going ever faster but bus standards singularly failing to keep up, the need for manufacturers to differentiate their machines by their bus technology is becoming ever more pressing. All indicators point to high-bandwidth, low-latency, highly-scalable system and device interconnects – or the big bus – will be that key differentiator. And, as reported […]
With processors going ever faster but bus standards singularly failing to keep up, the need for manufacturers to differentiate their machines by their bus technology is becoming ever more pressing. All indicators point to high-bandwidth, low-latency, highly-scalable system and device interconnects – or the big bus – will be that key differentiator. And, as reported (CI No 2,707), fault-tolerant systems builder Tandem Computers Inc last week unveiled an ambitious hardware-software big bus interconnect, which it calls ServerNet. Tandem says that it is simply not cost-effective to keep adding expensive processors and touts ServerNet as a System Area Network, or SAN, that enables any server component – CPU, storage or communications device – to interconnect directly without processor intervention. Instead of a processor on every data path, six-port, packet-switching routers implemented in an ASIC provide intelligent switching and wormhole routing with 300Mbps bandwidth per router (planned to double over time) and three microsecond message latency. Tandem says it means many of the large blocks of data required by text, image and video applications of the future will be retrieved through routers that will cost a few hundred bucks without having to pass through costly CPUs. Tandem says it had to start from scratch when creating its TNet/ServerNet interconnect after looking for an existing bus or network to fit its architectural requirement but finding nothing that would fit. Fibre Channel didn’t meet its requirement for short packet-switched transfers, although it performs well on long sequential transfers. Asynchronous Transfer Mode was deemed too costly given the large number of network attachments required in a system area network. Scalable Coherent Interface has a design centre for handling cache coherency, but costs and complexity exceeded Tandem’s limit. P1394 serial links wouldn’t meet its high-end requirement. The system area network it came up with has a 64 byte packet size, no memory-mapped input-output and is claimed to enable the on-line servicing of peripherals; isolation and containment of faults; a common input-output programming model; and comon input-output configuration and error management. TNet nodes can read and write portions of each other’s memories without requiring software execution at the remote node. Tandem said TNet can be used in various topologies – hypercubes, meshes and trees. ServerNet is an additional layer that logically and physically isolates processor buses from input-output buses (dispensing with memory-mapped input-output), and provides processor and bus interfaces for VME, PCI and other device inter connects. ServerNet links connect processing nodes either within a cabinet over the backplane, or to other cabinets that can up to 100 feet off, over copper cable. Tandem plans to use ServerNet to bring its still fundamentally different Integrity Unix and Himalaya NonStop architectures together into hybrid symmetric multiprocessing, clustered and massively parallel processing arrangements, which for the first time will use common storage and communications devices.
By William Fellows
ServerNet will appear by the end of this year on the first home-grown, fault-tolerant symmetric multiprocessing Unix servers in Tandem’s currently uniprocessor-only Integrity FT line. The one-to-four ways will use MIPS R4400 RISCs with a new symmetric multiprocessing Unix implementation and are likely to top out at eight processors. Its deal with Silicon Graphics Inc – Tandem rebadges symmetric multiprocessing Challenge boxes as Integrity NR servers – is likely to bite the dust thereafter. ServerNet’s second incarnation will feature in Windows NT servers, due in 1996, designed from the ground-up to use the next-generation MIPS R10000 part, also Tandem’s first R10000 implementation. Unix R10000 systems will follow. An issue still being worked through is how R10000’s integrated symmetric multiprocessing will dovetail with ServerNet. Both new symmetric multiprocessing lines will use a cohere
ncy bus in ServerNet to provide cache coherency between memory modules. The interconnect won’t find its way into to the company’s core NonStop business until late next year when it will feature in new high-end R4400 parallel nodes that may or may not appear under the Himalaya name. Although existing Tandem NonStop applications will be object code-compatible with its new systems, ServerNet – which replaces the company’s expensive TorusNet interconnect – cannot be retrofitted to any current products. As well as applications now being tailored for use over ServerNet, Tandem is also working with independent software vendors to implement under-the-hood techniques such as spotlighting, where existing database code is optimised for a select set of processors and devices. To do this without changing the original application it has to work what type of code changes can be made without having to present a separate application programming interface. It’s also going to have a bundled, integrated version of the Illustra Technologies object-relational database working in the guts of ServerNet that will lick relational and procedural data into objects that can be directed around the system over ServerNet. Tandem is hedging its future processor bet by positioning ServerNet as a RISC-independent interconnect, going so far as to say it has signed up MIPS until R10000 but has a clean sheet of paper in front of it thereafter. Moreover, having spent probably more than it should have done to develop ServerNet, the $2,000m a year company is looking to recoup some of its investment by offering ServerNet OEM. L M Ericsson Telefon AB has already agreed to take the next-generation Integrity FTs and there’s claimed to be one Japanese and one US OEM deal in hand. The danger of offering the product OEM is that someone might be tempted to create a better implementation than Tandem’s, although there’s little prospect of anyone beating its fault-tolerant expertise. Analysts say candidates could include Hewlett-Packard Co, which has already consigned the Exemplar systems it takes OEM from Convex Computer Corp to the technical world and whose T500s, are in their belief, looking a little long in the tooth. Other companies working towards ‘big bus’ interconnects include Sequent Computer Systems Inc, AT&T Global Information Solutions, Cray Research Inc, Silicon Graphics Inc, Pyramid Technology Corp, Data General Corp, IBM Corp and Compagnie des Machines Bull SA. Tandem will especially target ServerNet at IBM customers using SP2’s 40Mbps Vulcan switch.