Network Computingwww.networkcomputing.com

At the time of the test, only two models were available: the 12-slot Model 500 and the 10-slot Model 300. A six-slot expansion unit (Model 501) is available for the 500. A smaller version, the five-slot Model 200, is in the works. IBM has long-term plans to expand the line much as GDC has done with its line, while integrating routing and switching based on its 2216 and 2210 lines.

At the core of the 2220 lies IBM's Switch-on-a-Chip, which is also the central engine for the IBM 8260 Multiprotocol Intelligent Switching Hub. The Switch-on-a-Chip is a 16x16 switch that allows output queues to be built out of shared memory, and currently provid ing 400 Mbps per port.

Overall, the 2220's performance was in the average-to-good range, with excellent marks in frame relay latency. Standout features are its ability to optimize low-speed trunks by mixing cells and frames, its scalable NMS options and its tight configuration management via the NMS. Otherwise, it lagged the others in interface and service features. In particular, frame relay to ATM service interworking is not available, and the switch doesn't support ATM SVCs or SPVCs, nor frame relay SVCs.

Voice quality in our tests showed a slight degradation under high delay (greater than 300 ms) conditions. But in all fairness, this delay is outside the bounds for which IBM's echo cancellers were designed (32 ms); these echo cancellers are intended to compensate for delay introduced by four-wire to two-wire conversion through a PBX.

The 2220 line implements IBM's Networking BroadBand Services (NBBS) architecture, which describes both a logical view of th e software running in a node, and a des cription of overall network service capabilities. One special NBBS feature is the ability to mix ATM cells and variable-length frames on a single trunk. This can substantially reduce overhead and increase trunk utilization, which become particularly noticeable over slow links. Trunks can also be configured in ATM-compliant mode.

IBM provides good flexibility in applying services to interfaces, unlike Nortel's requirement to buy specific line cards for specific services. One Nways adapter can simultaneously run CE in both structured and unstructured modes, voice, HDLC and frame relay. The notable exception is in ATM service, where the line cards are specific to ATM.

An optional voice server adapter in IBM's product provides optimizations including compression, silence removal, idle removal, echo cancellation and fax detection. Voice services can be mapped onto channels residing on any card that bears the actual traffic, a more flexible configuration than Nortel's.

Unfortunately, the application of t hese adapters requires some careful analysis. Each voice server adapter handles fewer calls than a T1 circuit, supporting only 20 64-Kbps channels with compression and silence removal or with echo canceling. If you need all of these optimizations, the card supports only eight channels. Extension cards are available to drive this capacity as high as 56 channels with all optimizations.

Each NBBS link supports four logical queues and four levels of priority. Real Time 1 is a CBR queue for low delay/low jitter applications; Real Time 2 is a CBR or VBR-rt queue for applications that tolerate slightly more delay and jitter; Non-Real Time queue is for VBR-nrt traffic; and Non-Reserved queue is for VBR traffic that needs only best-effort delivery.

NBBS uses a proprietary algorithm called Control Point Spanning Tree to build an efficient multicast network for distribution of network topology information to all nodes, requiring only N-1 links to notify N nodes of a network status change. As for all switches in t his review, PNNI is not yet available for NBBS.

Each IBM switch is configured and managed by an Nways Administration Station (NAS), an OS/2-based PC running on the node. Hardened PCs that host the NAS, as well as modems for remote access, are a standard part of each switch. This PC stores configurations, error records and accounting information, and takes part in the initialization and loading of code into the switch. Unfortunately, this PC must be kept in working order at all times. If a switch reset is required (admittedly a rare occurrence), a dead PC means a dead node.