A good plan is essential when acquiring a multiprocessor system. You must know what the load will be and then buy more than you need to accommodate expansion. While we are not advocating over-buying capacity, this is an expensive investment, and you want to provision for expansion. It's never fun to revisit the vice president to ask for more money because the new system got overloaded in six months. Unless you look at your network in its entirety, no system or upgrade will be completely satisfactory.
Some simple factors play a role in your multiprocessor system purchase: politics, service, price, processor, bus architecture, performance and capacity.
So Sayeth the Sages
A strange political aspect of buying a new server is brand loyalty. An eternal argument ensues about the importance of brand conformity versus having best-of-breed solutions. If your main brand is the best solution, you'll avoid a huge mess. However, if you're a staunch Dell Computer Corp. shop and you discover that Silicon Graphics offers the perfect solution, should you switch? Moreover, should you switch completely or just in that instance? The relationship with a tried-and-true vendor is key, but maintain a healthy skepticism.
Familiarity with installation software, service procedures and technical support people at your vendor is a great boon. The danger lies in becoming so enamored with a vendor that you can't see the shortcomings. Switching vendors is easy when the products are shoddy but is a more difficult call when only customer service, repair, product selection or financial condition is poor. All these factors must be considered when choosing a vendor. Don't let your workstation purchasing decision unduly influence your server decisions. And because you purchased servers from a particular company, its workstations will not necessarily work better. If Ford Motor Co. bought an asphalt company, would Ford cars run better on Ford roads?
The server-vendor decision should be made separately from the decision over workstation vendor. Vendors will prattle on about "intangible benefits" of purchasing workstation and server hardware from the same vendor. If you can't define a benefit, it's not a benefit. Unquantifiable benefits may simply be T-shirts, CIO lunches and free golf passes. Let the vendor provide real pricing, warranty or service benefits.
I/O, I/O It's Off to Work We Go
By far, the most common bus architecture is PCI (Peripheral Component Interconnect). This standard was born at Intel and was confirmed and adopted in the early 1990s. The first motherboards to incorporate this standard were for Intel's 80486 processors. The specification rapidly grew to include both 32- and 64-bit data paths, and provide support for clock speeds of 33 MHz and 66 MHz. The 64-bit slot is longer than the 32-bit slot and is generally used for NICs and hard disk controllers. They greatly benefit from increased data rates with a 64-bit slot running at 66 MHz.
A 32-bit slot running at 33 MHz has a maximum data rate of 132 MB per second. A 64-bit slot running at 33 MHz has a maximum data rate of 264 MB per second. The 32-bit slot running at 66 MHz has a maximum data rate of 264 MB per second, while the 64-bit slot running at 66 MHz has a maximum data rate of 528 MB per second.
To run at the maximum speed, a PCI card must be designed for the higher speed and bandwidth. Older cards not designed to run at 66 MHz simply can't do so. The bus will detect the proper speed for the card and adjust accordingly. You can also insert most 64-bit cards into 32-bit slots if no 64-bit slots are available. The 64-bit cards will operate at 32 bits when inserted into 32-bit slots.
Hot-swappable PCI slots are also an important consideration. This feature is used to replace NICs, for example, without downing the server. Your OS must support this feature. Consult with your VAR or, even better, the system manufacturer about support for hot-swappable PCI slots. Complete support from both the hardware and the OS is necessary for this feature to work. Space and removal limitations in certain rack-mount environments may make hot-swappable PCI slots less important to you.
Memory Bandwidth
Several types of memory are on the market today. SDRAM (synchronous dynamic RAM) is the market leader, but Intel is aggressively pushing the Rambus memory standard. The Rambus standard has a higher theoretical performance in bandwidth-intensive applications, but there have been several problems with its adoption. The memory modules are difficult to manufacture, slowing the ramp-up of Rambus production. Technical issues with Intel's i820 (Camino) workstation chipset also scared manufacturers and further slowed the technology's adoption. But the biggest barrier to market adoption has been price. Rambus memory generally is almost twice as expensive as its SDRAM counterpart. The proprietary nature of Rambus technology accounts for a portion of its high cost: The company Rambus owns the patent for this memory and charges a royalty to manufacturers.
Another surging technology to keep an eye on is DDR (double data rate) SDRAM. This upgrade to the venerable SDRAM greatly improves memory bandwidth. It transmits data on the leading and falling edges of the clock cycle, effectively doubling the throughput. Most multiprocessor systems today support error-correcting SDRAM.
Processors Galore
Choosing a processor usually has more to do with your big-picture platform choice and software requirements than in finding the processor that's really best for you. If you're a Sun Microsystems shop, SPARC likely becomes your processor of choice. If you have some flexibility in choosing, Intel's line of x86 processors is inexpensive and solid. The Intel Pentium III Xeon processor and the Pentium III are other good options; the primary difference between them is the amount of L1 cache memory available. The Xeon has quite a bit more memory: It has up to 2 MB of L1 cache memory, whereas the Pentium III has only 512 KB.
If you use a RISC Unix variant, such as Compaq Computer Corp. Tru64, IBM AIX or Linux, your processor of choice will likely be Compaq's Alpha or Motorola/IBM's PowerPC. The Alpha can be quite expensive. The prime reasons to go with an Intel processor are price and OS choice. Microsoft Windows and Novell NetWare are pretty much Intel-only platforms.
The 64-bit processor market, which has been dominated by the Sun UltraSPARC and the Alpha processors, is moving toward the mainstream. Complexity is increasing and 64-bit processor technology is becoming affordable. The 32-bit processor market leader, Intel, is making a push into this space with its Itanium (see "Itanium: The EPIC New Processor," page 105).
Advanced Micro Devices (AMD) has a new 64-bit processor code-named Sledgehammer. AMD has taken a different route to 64-bit computing than has Intel. While Intel is proposing a radically new 64-bit architecture, AMD is extending the x86 architecture to 64 bits. This approach will help protect your investment in 32-bit software by letting it run at full speed while taking advantage of 64-bit software compiled for Sledgehammer. Compilers and code should also be easier for this reason. AMD expects to ship the Sledgehammer in the first half of 2002. These new chips from AMD and Intel have the potential to change the processor landscape.
Storage: Inboard or Outboard?
The number of drive bays your new server will support affects the form factor and height of the unit. If you're attaching your server to a SAN (storage area network), you don't have much need for internal storage. However, if the unit will be in a remote location without a SAN, you can purchase systems to support as many as eight externally accessible hot-swappable SCSI drives. External access to your drives is especially important in a rack-mount configuration. Units mounted in racks should be mounted on rails, especially units larger than 2U. Having panels to access hardware without removing the unit from the rack is also beneficial. Hardware with the greatest chance of failure should be accessible from the outside.
The power supply is another important factor. Any unit you purchase should have at least two hot-swappable power supplies. Each should be capable of running the entire system. Systems also should have hooks in your OS to alert the admin of noncritical failures, such as with fans.
Mr. Goodwrench on Call
The support program offered by your system manufacturer will have a profound effect on your satisfaction with the system. When you've determined the business significance and uptime reliability needs of a given system, carefully consider your repair options. Warranty length is a consideration, but it's not a primary one. The main considerations should be the location and nature of your service provider. If you purchase a service agreement, find out who will be fulfilling it. If it's a national service, like IBM, find out where the nearest representative is and what the minimum response time is. Make sure you're not just contracting the local VAR through a third party.
Next, find out how the supplier is stocked, since the location of replacement parts is also a concern. Make sure the parts with the highest failure rate are no more than four hours away. All other parts can be brought in on a 24-hour basis.
After you've made the preliminary decision on a machine, talk to manufacturer representatives. Ask them what part fails the most. Although it may take considerable effort, you have to get past "Our equipment is so reliable, it never breaks." Reassure the representative that you have a genuine concern about parts availability and replacement.
If your main point of service is your local VAR, make sure the company is up to the task. Technicians should have Comp/TIA A+ PC hardware certifications. A+ is a good starting point for a hardware technician. Additional certifications from the manufacturer are also essential. Make sure the VAR is willing to put a pager on a number of these qualified bodies and make them available around the clock. Your data and operations deserve no less.
Send your comments on this article to Steven J. Schuchart Jr. at sschuchart@nwc.com.
RSS
