|Feedback To Author | Feedback To Webmaster

Copyright © 1998 by Fabian Pascal. All Rights Reserved

 

32Bit Computing

By Fabian Pascal

Disk Subsystem

A previous article elsewhere (www.microtimes.com/179/32bit.html) discussed the basics of the disk subsystem and provided guidelines for choosing a disk controller and drives. It is recommended to review it as context for this article, which evaluates the following SCSI disk subsystem:

Controller: Adaptec 2940UW Dual

* SCSI: ultra-wide

* firmware: 1.33

* driver: 4.00.1411.01

Drive: Seagate Cheetah 9LP

* SCSI: ultra-wide

* firmware: 0004

* buffer: 1024KB

Drive: IBM Ultrastar 9ZX

* SCSI: ultra-wide

* firmware: 03B1

* buffer: 700KB

Drive: Quantum Viking II

* SCSI: ultra-wide

* firmware: 4110

* buffer: 512KB

Cable: Granite Digital SCSIVue

* SCSI: 68pin HD ultrawide Teflon Ribbon

* 68pin Gold HD Diagnostic Active Terminator

Note: The dual version of the Adaptec controller is sold only to OEMs. Users can purchase the 3940UW unit, which performs the same, but may be more expensive. An alternative is Diamond Multimedia's FirePort 40 dual controller, whose performance proved in tests the same as 2940's.

The test system is the currently recommended configuration(which is best-of-breed system as of each article and, thus, may include components tested, but not yet covered), together with descriptions of the testing/benchmarking tools, with links to their home pages and recommended software .

Single Drive Performance

The two drive specifications most commonly used as measures of drive performance are average access time and sustained transfer rate. As already explained, loosely speaking the former is more important for random access operations (e.g. databases) and the latter for sequential operations (e.g. graphics, or video applications). Table 1 compares the two actual rates to the manufacturer specs).

Table 1: Hard Drive Performance (HD Tach)

=================================================================
	                    Cheetah         9ZX          Viking2               
                        ----- ------  ----- ------  -----  ------
                        Specs Actual  Specs Actual  Specs  Actual
-----------------------------------------------------------------
Avg. Access Time (ms)    8.2  8.9     10.1   9.8    13.7   11.7 

Sust. Xfr Rate (MB/s)
 Max                    21.3  18.1    17.0  16.8    14.0   14.2
 Min                    14.5  11.5    10.5  10.2     8.0    7.8 
 Avg                          16.1          14.9           11.5
 CPU Util. (%/MBs)             .32           .34            .34  
=================================================================


Table 2: Single Drive Performance

=============================================
                   Cheetah     9ZX    Viking2     
---------------------------------------------
WinBench 98
---------------------------------------------
Business             2960      3017     2203    
High-End             6573      6840     5583 
=============================================
FAB
---------------------------------------------
Sust. read (MB/s)
 FAT                  8         8        8 
 NTFS                 6.6       6.6      6.6                 
Efficiency Mark
 FAT                 37.5      36.7     37.2 
 NTFS                40.9      40.9     40.9
=============================================
Winstone98
---------------------------------------------
Business             29.2      29.3     28.8   
High-End             32.1      32.2     32
=============================================
ThreadMark
---------------------------------------------
Xfr Rate (MB/s)      12.21     10.89     8.79 
CPU Util. (%/MB)      1.39      1.37     1.24  
=============================================

Table 3: Multi-Drive Performance

=============================================
ThreadMark
---------------------------------------------
Xfr Rate (MB/s)      12.21     10.89     8.79 
CPU Util. (%/MB)      1.39      1.37     1.24  
=============================================

* Hardware configuration (recommended system)

* Windows NT 4.0 Workstation, Service Pack 3, default caching;

* Display mode: 1280x1024x24bit @75Hz refresh

 

The actual access times are within the specs, except for the Viking, which exceeds them (product manager Skip Shapiro says the initially published specs were subsequently improved). The Cheetah has the shortest time; the 9ZX and Viking have 10% and 31.5% longer times, respectively.

The transfer rate varies across a drive, decreasing from the outside in, as shown in Figure 1 for the Cheetah and the Viking. This is why Table 1 gives three transfer rates: maximum (the most outer section of the drive), minimum (the most inner) and the average. Cheetah's rate drops by 36% from outer to inner sections, 9ZX's by 39% and Viking's by 44%. The 9ZX is close to its specs; the Cheetah is under them by 15% (max.) and 21% (min) and the Viking by 45%.

Figure 1: Transfer rates across drives (HD Tach)

 

 

Cheetah 9LP

 

Viking II

The Cheetah has the highest transfer rates. 9ZX's are lower by 7% (max.) and 12% (min.) and Viking's by 21% and 32% respectively.

CPU utilization is important in multitasking systems: the less the disk subsystem utilizes the CPU, the more the other simultaneous tasks will have available and may run faster. All three drives' (normalized) rates -- % of CPU time per MB/sec -- are similar.

Ziff-Davis' WinBench 98 benchmark exercises the disk subsystem in isolation, by throwing at it all the Winstone disk operations at once. Table 2 shows that under this high pressure the Ultrastar drive does slightly better than the Cheetah, with the 7,200rpm slower by about 26% (business) and 18% (high-end). Will these differences show up at the application level?

The sustained transfer rate reflects throughput of data delivery from the disk subsystem to the system board. FAB (File-System Access Benchmark measures theSustained Read Speed --how much data the disk subsystem is capable of delivering all the way to applications -- by taking into account the operating system and its file system overhead. Table 2 shows this rate to be much lower than the regular sustained rates and is actually identical for all three drives! The NTFS rate is lower by 18% than FAT.

FAB does lots of disk "hyper access" to exercise the disk subsystem and produces an Efficiency Mark that measures the subsystem's "horsepower" from a point of view closer to applications and users. According to FAB's developer, Vincent Giovannone, drives with similar Marks should feel similar to the user.

Real applications tend to mirror FAB results fairly closely. For example, throwing a disk into MS-DOS compatability mode, most benchmarks will show that it does not affect performance much, but FAB will take a giant hit. Accordingly, real-world application experience is that MS-DOS compatability mode is considerably slower."

If so, according to the results in Table 2, all three drives -- including the 7,200rpm Viking! -- will feel the same to users (here, NTFS scores are higher by about 8% than FAT, which Giovannone attributes to better NTFS caching). And indeed, Ziff-Davis Winstone 98 scores -- both business and high-end -- are practically the same for all three drives, validating Giovannone's claim.

Applications on stand-alone systems are single-threaded. Some performance differences do show up during multithreaded operations. Table 2 shows the Cheetah has the highest ThreadMark transfer rate, with the 9ZX and Viking 10.8% and 28% lower, respectively. ThreadMark is, however, somewhat controversial; also, it measures sequential operations and the multithreaded operations characteristic of servers tend to be random.

Multidrive Performance

Concurrent use of multiple drives is an issue mostly for servers and high-end video applications (e.g. video editing).

For the stand-alone user of business applications, the combined maximum sustained rates of two drives will not saturate the 40MB/s ultra-wide SCSI bus (e.g. 2x18.1=36.2 for two Ultrastars, 2x16.8=33.6 for two Cheetahs and 2x8.0=16.0 for two Vikings). Moreover, depending on the application, the actual overall rate is likely to be only slightly higher than one drive's rate, because as long as one drive has the bus, the other will have to wait. As Seagate's Bob Birk puts it:

"On one SCSI bus only two devices can transfer data between themselves at any given time. With one drive on the bus -- controller and drive -- speeds work out faster, because the drive does not have to wait for the bus. With two drives, the first one to gain access to the bus has the advantage. If the second drive has the read data collected into its buffer, but the bus is in use by the first drive, it has to wait, regardless of whether the two drives are of equal speed. There is no perfect interleave, as each drive is running at its allotted speed as an independent entity, and when it has data to transfer over the bus, the bus may or may not be in use. This wait time results in lower throughput overall."

In the absence of twin drives, Table 3 shows the ThreadMark rate for two 10,000rpm 9GB drives --the Cheetah and the 9ZX--at 20.31 MB/s. While the combined rate is 18.1+16.8=34.9, the actual rate is only about 2MB/s more than the Cheetah's and 3MB more than 9ZX.

 

Conclusion

Ultrawide SCSI controllers seem to perform similarly. Therefore, choice can be based on price and vendor reliability and support. A dual controller, which provides two SCSI buses, one wide and one narrow, is recommended, because it can keep slower, narrow devices such as removable or CD-ROM drives (to be covered next month) separate from the faster, wide hard drives, thus avoiding potential waits by the latter for the former.

The performance differences between 10,000rpm and 7,200rpm drives shown by the commonly published synthetic benchmarks (such as WinBench) are due to exercising the drives to the max. Operating system overhead and applications, however, make these differences imperceptible to users. Another way of putting it is that business applications on stand-alone systems are not demanding enough to take advantage of differences. Consequently, SCSI hard drives can also be chosen based on price and vendor reliability and support.

The rotational latency differences between 10,000 and 7,200 drives become important in multidrive server environments and/or more demanding applications such as video editing. Ultra2, Fiber Channel and RAID disk subsystem configurations should be considered for such environments.

(I would like to thank HPC's Vincent Giovannone, TestaCD's Greg Smith, Granite Digital's Tom O'Neill, Seagate's Bob Birk and Tyson Heyn, Quantum's Skip Shapiro, Adaptec's Chen Lu and Diamond's Dawn Tappy for their cooperation and help).

| Feedback To Author | Feedback ToWebmaster

Copyright © 1998 by Fabian Pascal. All Rights Reserved