Originally published in the January 1995 issue of Advanced Systems.


Hot boxes: Three ways to serve thee

We compare an Alphaserver 2100, an HP 755, and a four-Pentium ALR Revolution Q-4SMP.

By Mark Cappel

It's fun to watch computer makers scramble to catch what they hope will be bundles of money thrown about by organizations downsizing to midrange Unix computers. Those benefiting, of course, are users who have richer choices today in servers than ever before.

As part of our ongoing series of server reviews at Advanced Systems, the Advanced Systems Test Center (ASTC) gathered three very different Unix computers for this comparative review.

Our test subjects included a Hewlett-Packard 9000 700 model 755, a deskside computer using a single 99-MHz HP-PA7100 CPU. Our 755 happened to be configured with 192 megabytes of RAM and four hard drives, totaling 7.2 gigabytes of capacity. From Digital Equipment came an Alphaserver 2100 4/200 powered by a pair of 190-MHz Alpha CPUs (Digital says Advanced Systems is the first magazine to test a multiprocessing 2100). Our 2100, a deskside computer, sported 256 megabytes of RAM, a pair of 2-gigabyte disks, a floppy drive, and a CD-ROM drive. Finally, from Advanced Logic Research, we tested an ALR Revolution Q-4SMP a multiprocessing deskside computer with four 90-MHz Pentiums, 128 megabytes of RAM, 4 gigabytes of disk space in an on-board RAID-5 subsystem, a floppy drive, and a CD-ROM drive.

Our favorite? It depends on the application. For floating-point problems we give the nod to the Alphaserver 2100. For general-purpose use, the ALR SMP box offers acceptable performance and tremendous flexibility. The HP 755 can't keep up with its MP competitors in computing tasks, though it has good I/O, a trait we've noted before in HP's Unix line.

Digital Equipment Corp. Alphaserver 2100 4/200
In October 1993, the ASTC reviewed a deskside computer powered by an early 200-MHz Alpha chip. We wondered at that time what the computing landscape would be like if Digital could find a way to tie several Alpha chips together in an SMP configuration. Now, 15 months later, we know. Digital reports selling 6,000 Alphaserver 2100s since April 1994. Sun, by contrast, has sold 6,000 SPARCserver 1000s since its introduction in April 1993, a full year earlier. HP moved 10,000 HP 9000 model G and H systems (midrange Unix computers) in 1993.

Digital and its value-added resellers offer 190- and 275-MHz versions of the 2100. Both (like almost all of Digital's Alpha-chip product line) operate with Windows NT, VMS, and OSF/1. Our 2100 had OSF/1 3.0 pre-installed. While we didn't have time during the course of this review to port applications, the ASTC's previous experience porting to OSF/1 indicates the operating system is no more difficult to port to than other flavors of Unix.

Despite Digital's bragging rights to the PDP-7 being the place of origin for Unix back in the late 1960s and early 1970s, today Digital is considered a newcomer to the Unix business because of the freshness of its Alpha hardware and OSF/1. Combined with Digital's less-than-stunning desktop Unix sales, this translates into a relatively few 3,000 OSF/1 applications as compared with 8,000 for Sun, 5,000 for HP, and 18,000 for SCO. Keep in mind everyone counts Unix applications equally -- a public-domain front end to spell gets counted the same as an accounting package. Nonetheless, if you can find the applications you need for OSF/1, the 2100 is a great home for them.

Digital's OSF/1 staff needs to improve OSF/1's administration tools. While the ASTC has not used every Unix popular today, we can say OSF/1's admin tools are the weakest of the big vendors. If you're adopting OSF/1 we recommend that experienced administrators rely on their stash of trusty scripts and that admin greenhorns hightail it to the nearest community college for a basic Unix administration course.

Lately, Digital has made a big deal of its workstation and server line operating with the 64-bit Alpha chip. Compared to 32-bit chips, which every other major chip maker (except MIPS) sells today, a 64-bit chip offers much larger address space. However, Advanced Systems has so far been unable to find a commercial application that exceeds the address space offered by 32-bit chips, hence requiring a 64-bit chip. (If you know of one, send a message to mark.cappel@advanced.com.) While having the ability to run 64-bit applications isn't bad, by the time applications requiring 64-bit address space become common, users probably won't want to run them on 1995's machines, given the mayfly life span of computers today.

Don't let our kvetching about Digital's workstation marketing steer you away from the 2100. If you need to combine high-speed number crunching and general-purpose file service in a chassis with room for growth, look no further.

ALR Revolution Q-4SMP
If "open systems" means "you can run anything" you can't get much more open than the ALR Revolution Q-4SMP.

We tested the Revolution using The Santa Cruz Operation's Open Desktop 3.0 with MPX 3.0. MPX is SCO's multiprocessor extension to its SVR3-based Unix.

Like exotic sports cars, everything the handsomely styled Revolution has, it offers in almost reckless abundance. It boasts eleven 514-inch half-height drive bays and a pair of 312-inch bays, seven EISA slots, three VESA local bus slots, space for four 90-MHz or 100-MHz Pentiums, and up to one gigabyte of error-correcting RAM. We thought IBM went fan-happy in its RS/6000 590, but even that didn't prepare us for the Revolution. We counted 10 fans, but more may lurk in there.

But like exotic sports cars where reliability takes a metaphorical cramped backseat, the Revolution Q-4SMP challenged our patience at times in matters small and big. In the nit department, the Revolution's flimsy door covering access to the floppy drives and CD-ROM stopped closing after a few hours in the ASTC. Changing the Ethernet card from BNC to AUI required us to halt the machine and boot it in DOS. These piffling complaints we can forgive.

But in the serious problems department, Pentium No. 3 (ours had the maximum of four) died in the middle of a Business Benchmark run, hanging the computer. According to ALR, a sick CPU is supposed to cause the computer to reboot, whereupon SCO Unix and the hardware reconfigure the system to operate as before but with one less processor. Our Revolution froze, and we couldn't restart the computer until we replaced the dead daughtercard. ALR's technical support couldn't explain this aberrant behavior or the reason for the Pentium turning belly-up.

In the administration department, the SCO-ALR combination earned a mixed score. We cast a wary eye on a system that requires booting another operating system to change an Ethernet port's configuration. And we wonder why the ALR handled the dead Pentium so poorly. On the other hand, SCO Unix' administration is functional and (aside from the Ethernet port business) complete, including a section on setting up automatic backups. SCO's administration tool is curses based.

Despite our quality concerns, ALR's sharp pricing and wide-ranging operating systems choices make the Revolution Q-4SMP a compelling server.

Hewlett-Packard 9000 700 model 755
Hewlett-Packard offers slow and fast 755 models. We got the slow one, which HP's literature colorfully describes as the "popular 99-MHz version." Both it and the "enhanced 125-MHz version" use a single PA7100 CPU.

Our favorite feature on the 755 is a column of LED indicators that blink, respectively, when the machine is running normally (the "heartbeat" LED) with disk I/O, with network I/O, and with good AC power. Early in our testing, we had trouble getting the 755 to work and play well on our lab network, due to our particular 755 expecting to bind to another server. A glance from across the lab at the dark heartbeat LED was our first indication the 755 didn't like its new surroundings. Maybe we're just old-fashioned, but we wish all servers had rows of indicators, like the 755, that blinked with greater frequency as the system got busier so system administrators could get quick insights into the health of their servers without being near a console (or a user throwing a fit).

The 755 is the skinniest deskside of the three tested. While this allows the 755 to fit in tighter spaces, it also means the computer has less space for disk drives and add-in cards (see the table Servers at a glance). Like the 2100, the 755 appears ruggedly built with stout components.

Our 755 came with a dumb terminal, so HP's SAM (system administration software) wasn't showing its pretty side. In spite of this disadvantage, we rate SAM the best of the lot and one of the top Unix administration tools. Which is not to say SAM couldn't be improved. HP should add password aging to this tool, as well as automated backup and restoration features.

We tested each server with the SPEC CPU and compiler benchmark, and the Neal Nelson Business Benchmark 2.4. (In future server reviews, we'll add the SPEC_nfs [a.k.a. LADDIS] NFS benchmarks to our list of system tortures.)

In much the same way car makers touted their EPA highway mileage numbers during the 1970s and '80s, computer makers today bandy SPEC numbers in an attempt to impress computer buyers. SPECint92 consists of six freely available C programs that use integer calculations. SPECfp92 is made up of 14 Fortran and C programs that rely on floating-point calculations. (For a more detailed discussion of SPECint and SPECfp, send an e-mail message to spec92@advanced.com.) But SPECint92 and SPECfp92 depend just as much on clever compilers and tweaking as they do hot CPUs. This became very clear during our tests of these servers.

In early 1994, the vendors overseeing SPEC realized the spoofery was getting out of control and tried to contain themselves by using a single C and Fortran compiler and the same compiler switch settings for each of the 20 programs. Results from this more controlled environment SPEC dubbed SPECbase_int92 and SPECbase_fp92. Computer makers were supposed to start quoting these numbers last summer. So far, only Digital is disclosing SPECbase results routinely, and only Digital has provided the ASTC its SPECbase "wrappers" (a form of make file for running SPEC).

Digital uses the Kuck & Associates KAP preprocessor to generate the SPEC results it quotes in marketing materials. (KAP rearranges code prior to compilation to optimize performance.) Since the majority of developers do not use KAP, the ASTC usually does not run SPEC with KAP enabled. We ran SPEC and SPECbase with and without KAP enabled and will disclose our KAP-enabled SPEC numbers here to make the playing field level, since the other vendors didn't give us SPECbase wrappers.

Running SPEC on the HP was maddening. Our initial SPEC runs resulted in integer scores roughly a third of those HP reports. After checking with HP, we learned we were using HP's freebie C compiler. We ftp'ed HP's high-performance compiler and ran SPEC again. This time our results were about two-thirds HP's advertised numbers. Another consultation with HP revealed we did not have HP's Performance-Based Optimizer (PBO), a utility akin to KAP. After getting a copy of PBO, our results matched (and in one case slightly eclipsed) HP's results. The lesson for HP users is clear: For maximum performance, avoid HP's free compiler and get a copy of PBO.

SCO isn't often asked for help in running SPEC on its operating system, and due to the limited time we had with each machine, we were able to generate only integer results for the ALR Revolution. (ALR doesn't quote SPEC results for the Revolution.)

As you might expect with a pair of 190-MHz engines under the hood, the Alphaserver 2100 clobbered the four-Pentium ALR Revolution and HP 755 in compute-intensive tests. In the single-CPU SPECint92 and SPECfp92, the 2100 generated 133 and 163, respectively. The HP earned 108 and 166. The ALR posted 49 SPECint92.

SPEC has a version of SPECint and SPECfp that can work on multiprocessing computers. SPECrate_int and SPECrate_fp run several copies of SPECint and SPECfp and measure the amount of work completed in a period of time. The test scales fairly linearly with the addition of CPUs. For our tests, we ran one copy for each CPU in the server.

Here with two Alphas cranking, the Digital also scored well, generating 5,674 SPECrate_int92 and 7,275 SPECrate_fp92. The 99-MHz HP 755 earned 2,549 SPECrate_int92 (a result slightly higher than HP's official number) and 3,933 SPECrate_fp92. The ALR Revolution measured 3,548 SPECrate_int92, and we did not generate a SPECrate_fp92 result.

The HP 755 acquitted itself in our I/O tests. (See Analyzing the Business Benchmark 2.4 tests)

Documentation & support
We give a perfect score (5.0) to documentation that is so well written people would read it even though they didn't have the product. Perfect documentation is accurate, well indexed, and sports an installation checklist and tutorial. On-line docs should consist of man pages and fancy hyperlinked reproductions of the printed manuals.

Despite our requests, Digital declined to send us the full paper-documentation set for OSF/1, while instead referring us to the on-line documentation, which is delivered on a CD-ROM. The on-line docs are fine, though we'd prefer to have one set of paper docs around for each system in case of emergency.

We found HP's paper documents to be voluminous and complete, though we wish HP would organize topics better so a harried sysadmin doesn't need to search through so many volumes to complete a task. Fancy on-line docs were not installed on our 755, probably due the fact that it was saddled with a dumb terminal for a console.

We received SCO's documentation straight from Santa Cruz. It's fine.

We give a perfect score to accurate, timely support that is available 24 hours a day, 7 days a week, and free for the life of the product. Support should be available via telephone, fax, and e-mail.

Each of the vendors offer a range of hardware and support policies. Digital offers the best hardware warranty in the business -- three years on site, free of charge. ALR has a single SCO expert available to answer questions.

What to buy
If CPU performance is everything, the Alphaserver 2100 is the machine for you. It is expandable from one to four CPUs. And if the 190-MHz Alphas don't have enough number-crunching urge, Digital offers a faster 2100 using 275-MHz Alphas. After comparing the performance results for the 2100 to its price tag, it's easy to see why Digital sold 6,000 of these so quickly.

The HP 755 we tested offers good I/O performance. For our money, we'd rather consider the 125-MHz version of the 755, or consider a cheaper, more compact HP E55. Our experience shows programmers need to take care if they wish to wring the maximum performance from the 755.

For general-purpose use, the low-cost ALR Revolution Q-4SMP is a winner. We wouldn't hesitate to use the Revolution for general-purpose file and application service despite its weak CPU performance even with four Pentiums, resting assured we could switch operating systems if upper management decided to standardize on something other than Unix. We would, however, be sure to get a hardware support contract.

About the author
Mark Cappel (mark.cappel@advanced.com) is executive editor at Advanced Systems. Cedric Higgins, manager of the Advanced Systems Test Center, generated the SPEC and Business Benchmark results for the computers tested. For additional benchmark data, send an e-mail message to janserver95@advanced.com. To receive more information about SPEC, send a message to spec92@advanced.com.

[Copyright 1995 Web Publishing Inc.]

If you have problems with this magazine, contact webmaster@sunworld.com
URL: http://www.sunworld.com/asm-01-1995/asm-01-server.html.
Last updated: 1 January 1995.