ry support are all enabling these systems to provide performance similar to that of departmental servers, but at prices closer to those of desktop PCs.
"It's hoped that all these efforts will push NT systems deeper into corporate enterprise networks," says Tony Chen, director of the PC department at Amaquest Computer (Taipei, Taiwan).
Although server-class processors developed by Intel offer more speed, overall system performance is still hampered by I/O bottlenecks because these chips are not specifically geared for handling device interrupts. I/O interrupts occur at a server when a communication request is issued by, say, a client/server transaction application. Depending on how the I/O device is implemented, it can interrupt the
CPU once or many times during a single I/O operation.
The I
2
O architecture, due to its ability to diminish these network bottlenecks, is viewed as a key technology for boosting the performance of PC servers. With I
2
O, interrupt-intensive I/O processing is off-loaded from the host/server processor onto I/O devices, allowing the host CPU to concentrate on its primary job, data retrieval. Vendors say that high-speed I/O performance is important for a growing class of I/O-intensive applications, such as video, groupware, and large databases.
In addition, I
2
O puts an intelligent device between a peripheral and the OS. Therefore, peripheral vendors need to write only one driver to the I
2
O standard; the OS then communicates with the I
2
O subsystem.
Network servers that support both the Intel 960 RISC chip and the I
2
O architecture are expected to appear in volume this year. Many server vendors, including those from Taiwan, began showi
ng their first full system implementations of the I
2
O architecture at Comdex in November. During that show, three Taiwan companies -- Acer, Amaquest, and Tatung -- unveiled I
2
O-compliant systems, and Asustek Computer demonstrated its I
2
O server motherboard. "I
2
O will give Intel and Microsoft their first real shot at competing head-to-head with Unix servers," says Chen.
Hardware integration is not the only factor that's important to the success of I
2
O; full support from Windows NT is a bigger issue, according to Cliff Chou, director of the networking systems business unit at Acer. "Giving the OS the ability to take advantage of it is the real challenge," he explains. "NT isn't there yet." Microsoft isn't expected to offer full I
2
O support until Windows NT 5.0 ships sometime this year, Chou adds.
Currently, Microsoft offers I
2
O support for Windows NT 4.0 in the form of a Service Pack. Several server vendors in Taiwan have b
een working with an early version of I
2
O running on NT 4.0 since late 1997. In addition, Novell, together with IntranetWare, endorses the I
2
O standard and NetWare 4.11, which should be available early this year. Novell has also promised to build I
2
O support into its next server OS. SCO's Gemini NetWare supports the Intel architecture as well.
The big question, of course, is: How much of a performance gain can users expect? The short answer is that it depends on the application, but I/O-intensive databases could see performance gains of 60 percent or more, according to server vendors. Testing is done by server makers to see how much of the I/O work load the i960 processor can handle.
Ideally, the new I/O architecture will significantly improve overall performance by having the i960 processor be responsible for all the I/O processing. But the initial rollout of the technology is not likely to achieve that goal, according to Jerry Lee, Asustek's associate vice preside
nt for product marketing and technical support. Some benchmarks show an improvement of around 10 percent to 15 percent.
Fault-Tolerant Designs
PC server vendors are responding to user demand by moving some degree of fault tolerance into their low-end and midrange systems. A host of new NT servers now includes error-correction-code (ECC) memory and an optional RAID controller, both of which are unusual in entry-level systems.
Chen says that servers are now being used more in distributed environments, where more fault-tolerance features are required. Downtime costs, coupled with data loss, are just too pricey for small-to-medium-size businesses. Equipped with RAID and "hot plug" technology, fault-tolerant servers can continue operating if any of their processors or other parts fail. Users can simply swap out the drives, power supplies, fans, and other parts while the server is still working.
Another quality that makes these Intel servers suited to client/server applications is their u
se of the Highly Parallel System Architecture (HPSA), which enables computer makers to incorporate two independent PCI "buses" -- one of which is then bridged to an EISA bus for support of older peripherals -- to eliminate data bottlenecks between peripherals, such as hard drives and the system's memory and processor. Hence, these systems include at least five -- and up to eight -- PCI slots.
Many servers devote a significant portion of their resources to the maintenance of high availability in their disk arrays. Most servers now come with two separate SCSI processors, a mirrored cache, battery backup, and a redundant power plant. This approach eliminates most single points of failure in the disk channel. In addition, Acer, Asustek, and Tatung have all included in their server solutions environment-monitoring sensors that monitor temperature, voltage, and fan speed; prevent crashes; and detect intrusion.
Asustek's
P/I-P65UP8 motherboard, for example, provides server manufactu
rers with seven PCI slots, as well as Symbios 53C876 and AHA-7880 PCI Ultra Wide SCSI controllers, plus the hardware required to implement I
2
O functions. The Asus SMP server motherboard will be targeted at Internet/intranet applications and database-server markets. Asustek offers a wide selection of CPU daughterboards that contain one or two Pentium Pro or Pentium II processors. A quad-processor daughterboard will be available by June. Other features of the Asus server motherboard will include as much as 1 GB of extended data out (EDO) DRAM, ECC memory for reliability and data security, and redundant hot-swappable power supplies.
Remote management is an essential component of a workgroup server. A good management program lets an IS manager monitor server status from a remote office. Intel's LANDesk, which is included with many workgroup servers, is a powerful utility that monitors a server's vital signs, such as the power supply's output voltages and the temperature inside the chassis. Some c
ompanies also offer proprietary vendor-developed solutions, such as the AcerAltos 9100 Advanced Server Manager (ASM) Pro and Remote Diagnostics Manager, and Amaquest's SNMP remote-control software.
True Clustering
Clustering is a technology for connecting separate computers, each running its own copy of an OS. For example, two separate four-processor Pentium Pro servers, each running its own copy of NT, can be connected to form a cluster. Currently, clustering is used primarily for high-availability and failure purposes, so vendors can offer a degree of fault tolerance in Pentium Pro servers.
Clustering essentially creates a quick-response, intelligent backup system in case one of the servers, or one of its components, fails. As a result, users are subjected to far fewer network outages. Until recently, clustering has generally been available only as a proprietary technology on Unix platforms.
Microsoft's Server Cluster server brings this technology to the growing NT server market
by providing computer vendors, resellers, and IT managers with the means to tie systems together in a fail-safe bond. Late last year, Microsoft released Windows NT Server 4.0 Enterprise Edition, which includes Cluster Server (MSCS).
Previously code-named Wolfpack, MSCS is a part of the set of APIs currently being developed to turn Windows NT Server and the BackOffice suite of applications into a clustering platform. Microsoft has said that a two-node variant, which will address failover fault tolerance, will be the first to arrive. Still another form of clustering at a higher level is currently being explored by Microsoft.
One of the first clustering servers to be developed in Taiwan is Amaquest's VPro "Cluster-in-a-Box" system. The server includes two separate 200-MHz Pentium Pro or 233-/266-MHz Pentium II system boards connected via 10-/100-Mbps Ethernet. The Ultra Wide SCSI controllers are configured as an independent bus, with each system's SCSI bus terminating at the storage processor in the
disk subsystem.
The VPro system can run in shared-bus mode. Each system can function independently of the other and can run all standard NT applications. This configuration provides service to one clustered system's connection without interruption to the other clustered system and prevents a total failure if a SCSI cable comes off and breaks the SCSI bus.
Amaquest plans to introduce four-way servers that use the Intel 440NX chip set. By the second quarter of this year, the company also plans to offer a new clustering server that contains three nodes in a box running Windows NT 5.0.
Where to Find
Acer, Inc.
Taipei Hsien, Taiwan
Phone: +886 2 696 1234
Fax: +886 2 696 1777
E-mail: Cliffchou@acer.com.tw
Internet: http://www.acer.com
Amaquest Computer Corp.
Taipei Hsien, Taiwan
Phone: +886 2 698 9366
Fax: +886 2 698 9376
E-mail: tony@amaquest.com.tw
Internet: http://www.amaquest.com.tw
I
2
O's Promise
