Tricord: A Mainframe’s Little Sibling

The Tricord Enterprise Servers are mighty big boxes. They're so big that, for shipping, they require specially designed crates with padded feet.

In each server's almost 4' tall cabinet, you can cram eight CPUs, up to 3.25GB of main memory, and nine half-height drives. The interior is shaped to duct air past critical high-heat components, and three redundant power supplies with 12-volt fans in each module pull air through. Still, wherever you set up the Tricord server, make sure you have adequate cooling, because these units generate a lot of heat. A systems engineer accompanies each new system to help set up and make sure the environment can handle the server.

Big Box, Big Features
The Tricord PowerFrame is not your typical PC server. It almost classifies as a mainframe or minicomputer--only it consists of standard PC and Intel components.

With starting prices of $70,000 for these machines, Tricord does not aim at the small workgroup server market. Instead, typical installations are for enterprise and mission-critical applications (such as Internet Service Providers--ISPs), where fault tolerance and availability are key issues. As a reflection of this focus, Windows NT is the operating system of choice for more than 20% of Tricord's total new system sales.

The PowerFrame can hold up to eight 166MHz Pentium CPUs (NT 3.51 and 4.0 both scale up to four CPUs right out of the box, but an OEM version of NT is necessary if you want to run five to eight CPUs), each with its own 2MB Level 2 cache module. The CPU-to-memory bus is part of the main 64-bit/ 33MHz system bus, which Tricord has dubbed the PowerBus. The PowerBus can reach 267MB-per-second (MBps) transfer rates. This bus also couples the CPU and Cache Subsystem (CCS, which uses its own 64-bit/66MHz internal data bus) to the other major system components. Each board holds two CPUs, and a system can have up to four CCS modules. Of the 10 total slots on the PowerFrame I reviewed, five are taken by the Intelligent Communications Subsystem (the ATM PowerLink, an optional 155Mbits-per-second--Mbps--fiber-optic ATM interface with its own processing hardware), the Intelligent Storage Subsystem (ISS, multichannel SCSI buses), the Main Memory Subsystem (MMS), the PCI Bridge Subsystem (PBS, two PCI buses, one EISA bus, one peripheral bus), and the Intelligent Management Subsystem (IMS, hardware monitoring and management). This architecture lets each subsystem arbitrate its bus usage via the System Bus Arbiter, with bus locking by the active device. This approach avoids wasting valuable time polling other subsystems.

All components except the ATM PowerLink are standard. A base configuration includes one CCS, one ISS board, the MMS (with the amount of memory you need), the PBS, and the IMS. Each CCS board (you can have up to four) holds one or two 166MHz Pentium CPUs with cache. The necessary subsystems take five slots, and you can fill five other slots in any combination of CCS and ISS boards. The CCS is designed for NT's symmetrical multiprocessing (SMP--although the CCS also supports asymmetrical multiprocessing for other OSs). The large, high-speed Static RAM (SRAM) Level 2 cache dedicated to each CPU significantly boosts the performance of multithreaded and SMP-enabled, 32-bit NT applications such as databases (SQL Server, Oracle Server) and messaging (Exchange, Notes). The PowerFrame's ability to use up to eight CPUs makes it a strong contender in enterprise NT environments where scaleability to high-end applications and heavy user loads are important.

The ISS is Tricord's primary claim to fame, and the PowerFrame can support up to six of these four-channel, fast-and-wide-and-deep, differential SCSI-2 controllers (two-channel versions are also available). Deep SCSI uses four bits for device addressing instead of three, so one ISS controller card can address up to 60 drives, with 15 drives on each channel (the last address is reserved for the controller). One PowerFrame server can support up to 201 drives, so you can have up to 2.1TB of disk storage with today's technology! The ISS is hardware accelerated for RAID levels 0, 1, 4, 5, and 10 (mirrored stripe sets, a.k.a. RAID 6), with disk hot-sparing (the SCSI backplane supports disk hot-swapping) and background disk rebuilding after failure and replacement. (For more on performance under different RAID levels, see the sidebar, "RAID Performance and NT: Configuring an Enterprise Server.") The ISS board includes a battery-backed 8MB controller cache, so if system power fails, you don't lose crucial data, even if it wasn't written to disk. Also, if the controller fails and the system goes offline, you can remove the card, transfer the cache module to a new controller card, reinstall it, and restart the system to begin recovery procedures. You can be confident that data not already flushed from the cache out to disk was not lost (this capability doesn't include data in main memory).

The PowerFrame's MMS can address up to 3.25GB of main Error-Correcting Code (ECC) DRAM memory. Unfortunately, Tricord chose to build proprietary memory modules rather than use off-the-shelf components. This approach raises the price of upgrading your system but has the advantage of improving performance with larger SIMM capacities than the industry standard--a 256MB single module is on the way. The MMS uses a 128-bit data path in parallel interleave (one 32-bit word from each module in a four-SIMM bank), and you can mix SIMM sizes if each bank is consistent (upgrade in groups of four).

The PBS serves as the conduit for all other peripherals such as network interface cards, modems, video, and keyboard. The eight physical expansion slots support any combination of PCI or EISA cards. All slots are bus-master capable. The PBS has dual-PCI buses (32-bit, 33MHz for 132MBps) bridged to a 32-bit EISA bus (33MBps). An integrated SVGA graphics adapter is incorporated on the PCI bus (1MB VRAM) with an 8-bit I/O peripheral bus for parallel, serial, floppy, and keyboard/ mouse. The PBS fully supports plug-and-play PCI and ISA cards for when NT has the appropriate drivers. But be aware that placing old and slow 8-bit or 16-bit cards on the high-speed PCI or EISA buses can drag down system performance significantly. Because the system has to service interrupts and transfer data at the speed of the slowest component on the bus, the system reduces the speed of all cards to match: Putting an old NE2000 NIC in your system will hurt your overall system performance. The moral of the story is to use cards designed for the task instead of using whatever's lying around (and if you're spending $100,000 on a server, why skimp on a $300 network card?).

Finally, the IMS offers everything from gathering system bus usage statistics to remotely managing server reboots. The IMS's two components are the board in the server (the Intelligent Management Processor--IMP) and the management software. The board is NiCad battery backed for up to 85 hours of server downtime. During such downtime, the board can dial out for administrator alerts, store critical last-error information, report the server's last-known and current status, and support dial-in management. The board has a built-in 14.4Kbits-per-second (Kbps) modem, 512KB of SRAM holding CPU code and data, 8KB of configuration data memory, 256KB of flash memory for runtime firmware and boot code, and connections to power, cooling, and temperature sensing systems. Because the IMS board sits on the main PowerBus, it can monitor system activity such as bus utilization and disk and processor usage.

The IMS management software (as you see in Screen 1) can run on the server or on a remote station (Windows 3.11 or later) via direct LAN connection or dial-up and can access any Tricord server on your network. The IMS management software can monitor power and cooling status for the main system and any attached disk-storage cabinets; show status, activity, and firmware revisions of all installed hardware subsystems (including LED displays, switch settings, and control codes); perform local or remote firmware updates to the IMP, PBS, or BIOS while the system is running; force disk recovery operations; perform extensive statistics analysis of all subsystems; reboot and reconfigure the system (load drivers, run EISA configuration utilities, and switch to a redundant CPU if available); access the system Event Log; and capture keyboard input and video output so that all administrator operations from a remote console are run on the server as if you were sitting at the server console.

The IMS will automatically handle as much fault prevention and recovery as it can, post such events to a system log, and notify the administrator when necessary or directed. The software can generate and send (via pager or Simple Network Management Protocol--SNMP) alerts on administrator-defined events, and follow rules for what to do in the event of a specific error or hardware failure. Hardware-level password security protects the server, so you don't have to worry about somebody dialing in to your server with the software and messing up your system.

Availability and Fault Tolerance
The Tricord PowerFrame is definitely for enterprise deployment, in both functionality and price. You have to pay for advanced features, but this system has plenty of them.

One crucial feature on an enterprise system is uptime. If you're betting your entire business on the idea that your server will always be available, it had better be! Tricord uses redundant controllers (SCSI, network, etc.), redundant power supplies, hot-swap disks, and so on to address this requirement by eliminating as many single points of failure as possible. So far, the company reports achieving 99.98% server uptime, or less than two hours of downtime in a year.

The PowerFrame's disk subsystem is fully fault-tolerant, with hot-spares, hot-swaps, duplexed controllers, and many levels of RAID. The memory subsystem can catch single- and double-bit errors, and correct single-bit errors before they cause problems. The server has three 500-watt power supplies. Any power supply can fail, and you can hot-replace it without affecting the system. Sensors protect the system from heat. They have thresholds that you can configure to automatically shut down the server if the sensors detect heat problems. And, you can configure a multiprocessor system so that if one CPU fails, the server will automatically reboot without it. You will be able to load balance and set up fault tolerance with multiple NICs, when NT supports these capabilities.

Is Bigger Better?
Although the PowerFrame is an enormous box for a PC server, it is a well thought-out, modular design: All components are easy to access and replace. The box has plenty of room for cards, additional drives, CPU and memory upgrades, and advanced networking options. Windows NT Magazine Lab tests (which I explain in, "Optimizing Exchange to Scale on NT.") showed that the hardware handles NT just fine, with CPU and memory scaleability that you expect from such a system.

No, this box is not for the meek. But Tricord is for those who have enterprise applications or heavy user loads to address with one server (to reduce administration hassles); want to consolidate an existing multiserver setup into one; or just want an extremely powerful, robust, and scaleable platform. Support plans are available in addition to the one-year onsite warranty you get with the initial purchase, and 24 X 7 technical assistance is always available over the phone.

PowerFrame Enterprise Server/166

System Configuration:Eight 166MHz Pentium CPUs; 1024MB of RAM; Seventeen 2.1GB Seagate disks; 4-channel ISS Controller; IDE 4X CD-ROM; One external storage cabinet Tricord Systems * 612-557-9005 Web: www.tricord.com Price: $216,055 (with four CPUs)