Novell is now a part of Micro Focus

Break the Web Server Speed Limit with a Web Server Accelerator

Articles and Tips: article

RON LEE
Senior Research Engineer
Novell Advanced Development Group

01 Aug 1998


Independent test results show how BorderManager Web server acceleration on a Compaq server beat out competing configurations costing thousands more.

Introduction

With the increasing popularity of Web servers as a means to dispense information both internally and externally, many network designers are looking for easy and inexpensive ways to boost the performance of NetWare-, Windows NT-, and UNIX-based Web servers.

Compaq and Hewlett-Packard recently used Netscape's Enterprise Server for NetWare and Novell's BorderManager Web server acceleration on Intel platforms to produce the fastest uniprocessor SPECweb96 benchmark results ever submitted to the Standard Performance Evaluation Corporation (SPEC).

This AppNote summarizes these impressive benchmark results and details the hardware and software configurations used in the tests.

For more information on BorderManager Web server acceleration, see the following related AppNotes:

  • "A Quick Guide to Web Server Acceleration", Novell AppNotes, Oct 1997

  • "Three Ways to Deliver Cached Performance to Your Internet and Intranet Users", Novell AppNotes, Sept 1997

  • "Web Server Acceleration with Novell's BorderManager: A Case Study of www.novell.com", Novell AppNotes, Aug 1997

For more information about the BorderManager family of products, see

http://www.novell.com/products/bordermanager/

SPECweb96 Benchmark Results

SPECweb96 (www.spec.org) is a Web server benchmark that measures the maximum number of operations (hits) per second a Web server can service. Over the past several years, this benchmark has been used to test a variety of Web servers on both Intel and non-Intel platforms. Figure 1 compares the Compaq and HP results to the rest of the SPECweb96 results.

Figure 1: Summary of SPECweb96 uniprocessor results (sorted in result order).


Date
HW Vendor
HW Model
Processor
MHz
RAM (MB)
MTU
Results

May 98

HP

NetServer LH 3/400

Pentium II

400

1024

1500

2131

Apr 98

DEC

AlphaServer 1200 5/533

Alpha

533

1536

4352

2045

Mar 98

Compaq

ProLiant 3000

Pentium II

300

512

1500

1639

Oct 97

DEC

AlphaServer 1200 5/533

Alpha

533

1280

4352

1631

Oct 97

DEC

AlphaServer 4000 5/533

Alpha

533

2048

4352

1586

Jul 98

DEC

AlphaServer 800 5/500

Alpha

500

1024

4352

1545

Oct 97

HP

9000/C240

PA-RISC 8000

236

1280

9218

1493

Jan 98

SGI

Origin 200 QC

MIPS R1000

180

1536

1500

1380

Apr 98

HP

NetServer LH 3/400

Pentium II

400

1024

9218

1342

Jul 98

HP

NetServer LC 3/350

Pentium II

350

1024

4352

1289

Feb 97

DEC

AlphaServer 4000 5/400

Alpha

466

1024

4352

1220

Oct 97

DEC

AlphaServer 800 5/500

Alpha

500

1024

4352

1158

Nov 96

DEC

AlphaServer 4000 5/400

Alpha

400

1024

4352

1157

Mar 98

Micron

NetFRAME 3100

Pentium II

400

1024

1500

1150

Jun 97

SGI

Origin 200

MIPS R1000

180

1024

4352

1062

Apr 98

HP

NetServer LH II/300

Pentium II

300

512

4352

1001

Aug 97

Sun

Ultra Enterprise 450

UltraSPARC II

296

1024

9180

884

Mar 98

Sun

Ultra Axi

UltraSPARC Iii

333

1024

9180

856

Aug 96

HP

9000 Model K460

PA-RISC 8000

180

2048

4352

801

Nov 96

DEC

AlphaServer 1000A 5/400

Alpha

400

512

4352

790

Mar 97

HP

NetServer 6/200 LH Pro

Pentium Pro

200

1024

4352

750

Jun 97

Compaq

ProLiant 800

Pentium Pro

200

512

4352

573

Oct 96

DEC

AlphaServer 2000 5/300

Alpha

300

640

4352

570

Feb 98

Data General

SiteStak TW5000

Pentium Pro

200

128

1500

500

Nov 96

HP

9000 Model D330

PA7300LC

160

512

9218

500

Sep 96

IBM

RS/6000 43P-140

PowerPC

200

448

4352

459

Jan 97

HP

Visualize B132L

PA-RISC

132

384

1500

430

Nov 96

HP

9000 Model D320

PA7300LC

132

512

9218

426

Nov 96

DEC

AlphaServer 300

Alpha

266

256

4352

421

Oct 96

HP

Visualize B160L

PA-RISC 7300LC

160

384

1500

380

Nov 96

IBM

RS/6000 43P-140

PowerPC

200

448

4352

330

Sep 96

Sun

Netra i 1/170

UltraSPARC

167

512

1500

310

Dec 96

HP

9000 Model D310

PA-RISC

100

512

9218

281

Jun 96

DEC

AlphaServer 1000A 4/266

Alpha

266

128

4352

252

Jun 96

HP

9000 Model D210

PA7100LC

100

416

4352

216

All of the results shown above are from Web servers with a single processor ranging from 200MHz Intel Pentium Pro processors to 533MHz DEC Alpha processors. The Compaq and HP results at the top of the chart (highlighted in bold) are among the fastest ever published, while their solutions are among the least expensive.

In March 1998, Compaq used a $19,000 ProLiant 3000 with a single 300MHz Intel Pentium II processor to produce 1639 SPECweb96 operations per second. At the time it was submitted, the Compaq result was the best of all uniprocessor solutions. In fact, it beat many of the dual-processor SPECweb96 solutions--including a $44,000 dual-processor Sun Ultra Enterprise 450.

In May 1998, HP used a $22,000 NetServer LH 3/400 with a single 400MHz Intel Pentium II Deschutes Slot-1 processor to generate 2132 operations per second. Once again, a Novell-assisted solution using BorderManager to accelerate the Netscape Enterprise Server for NetWare ran away with top honors. This solution even beat a $36,000 533MHz DEC AlphaServer 1200.

These SPECweb96 results are outstanding for several reasons:

  • Faster than a DEC Alpha. Both systems are faster than almost all other solutions, including several running on 533MHz DEC Alpha processors.

  • Faster than a Dual-Processor Sun Enterprise 450. Both systems beat the dual-processor Sun solution (see http://www.spec.org/osg/web96/results/ res97q3/web96-970819-02022.html for details).

  • Realistic RAM configuration. The Compaq/FastCache solution used less memory (only 512MB) than all other solutions at the high end, while other NT- and Unix-based systems required more than twice the RAM (1,280MB to 2,048MB) to compete.

  • Realistic network configuration. Both systems used Ethernet and a realistic maximum transmission unit (MTU) of 1514 bytes between the test clients and server. Nearly all of the other vendors' results used much larger, non-Ethernet MTUs, including a 4352-byte FDDI MTU and a 9218-byte ATM MTU. Results based on these MTUs aren't realistic and can't be applied to real-world intranet and Internet environments. In fact, a study produced by HP demonstrates that benchmarks configured with large MTUs can inflate benchmark results by nearly two times. The study found that when using a 576-byte MTU as a baseline for Internet communications, an FDDI MTU inflated results by almost 70%, while an ATM MTU inflated the results by nearly 200%.

  • Best price-performance ratio. The Compaq/FastCache solution's street price is approximately $19,000, while a slower dual-processor Sun Ultra Enterprise 450 system sold for over $44,000.

The bottom line is that, for corporate intranets, commercial content publishers, and ISPs, Novell BorderManager solutions serve more users faster and more efficiently--at a fraction of the cost of solutions based on UNIX and non-Intel processors.

How to Break the Speed Limit

The Novell solution used by Compaq and HP is unique because Novell's BorderManager and Netscape's Enterprise Server for NetWare coexist on the same server. This configuration yields the highest performance and the best price/performance ratio.

For this solution, the Web server operates on a secondary IP address within the server, while the Web server accelerator receives all requests for the Web server on the server's primary IP address (see Figure 2).

Figure 2: A Web server and Web server accelerator combined in a single server.

This combination of Web server and Web server accelerator is ideal for sites where server hardware is limited.

Detailed Configurations

The following sections detail the configuration and tuning parameters used for the Compaq and HP BorderManager solutions.

Compaq Configuration and Tuning Parameters


Hardware

Vendor

Compaq

Model

ProLiant 3000

Processor

1 Pentium II (300MHz)

Primary Cache

16KBI + 16KBD on chip

Secondary Cache

512KB (I+D) on chip

Other Cache

None

Memory

512MB

Disk Subsystem

1 4.3GB Seagate Barracuda, 6 4.3GB Seagate Cheetahs

Disk Controllers

Integrated SCSI-3 controller

Other Hardware

None


Software

Operating System

IntranetWare 4.11 with Support Pack 4B (INWSP4B.EXE)

File System Type

NetWare

SYS volume

OS, WS, and logging ran on a standard 4.3GB volume

Other Software

None


HTTP Software

Vendor

Novell

HTTP Software

Netscape Enterprise Server 3.5 for NetWare

No. of threads

1

Web server

Installed with defaults - loaded with NVXWEBUP.NCF (excludes the admin server)

Server Cache

Novell BorderManager 2.1 Web Server Accelerator with Support Pack 1A (BMSP1A.EXE) and Proxy Update 103B (BMP102B.EXE)

Log Mode

Standard


BorderManager

Configured to accelerate Netscape Enterprise Server for NetWare

Added NWAdmin for NT Snap-ins for BorderManager (BMNWNTB.EXE)

Updated MSM.NLM with 12-2-97 version

Updated ETHERTSM with 12-8-97 version

Updated PROXY.NLM with 3-2-98 version

Updated TCPIP.NLM with 3-12-98 version

Added FATFREFX.NLM

Added CACHEPER.NLM

Added LOCKPERF.NLM


Cache Volume

NetWare volume spanning six drives, each with a single 204MB partition

8KB volume block size, no suballocation, no compression


Benchmark Configuration

Maximum Requested Load

1650

Maximum File Set Size

888 MB


Network

Network Controllers

5 Intel EtherExpress PRO/100 Server Adapters

No. of Nets

5, with four clients per net

Type of Nets

Ethernet

Network Speed

100 Mbps

MSL (sec)

30 (Non RFC1122)

Time-Wait (sec)

60 (Non RFC1122)

Notes: Each net connected by a Bay Networks Baystack 100 Base-T hub. The Web server and controller are located on the fifth net. All nets at half duplex.

STARTUP.NCF File pmload FATFREFX.NLM pmload CACHEPER.NLM load LOCKPERF.NLM set maximum physical receive packet size = 1514 set maximum packet receive buffers = 2000 set minimum packet receive buffers = 500 set maximum service processes = 100

AUTOEXEC.NCF File set maximum concurrent disk cache writes = 500 set directory cache allocation wait time = 0.1 sec set directory cache buffer nonreferenced delay = 60 min set minimum directory cache buffers = 10000 set maximum service processes = 100 set upgrade low priority threads = off set maximum file locks = 100000

HP Configuration and Tuning Parameters


Hardware

Vendor

Hewlett-Packard

Model

NetServer LH 3/400

Processor

1 Pentium II (400 MHz)

Primary Cache

16KBI + 16KBD on chip

Secondary Cache

512KB(I+D) off chip

Other Cache

None

Memory

1.0 GB

Disk Subsystem

5 9GB 10K RPM Drives

Disk Controllers

Integrated Ultra-SCSI, I2O RAID


Software

Operating System

IntranetWare 4.11 with Support Pack 5B

File System Type

NetWare

SYS volume

OS and logging ran on a standard NetWare 4GB volume

Other Software

None


HTTP Software

Vendor

Novell

HTTP Software

Netscape Enterprise Server 3.5 for NetWare

No. of Threads

1

Web Server

Installed with defaults; loaded with NVXWEBUP.NCF (excludes the admin server)

WebPage Volume

Web server root on a NetWare 2GB volume

Server Cache

Novell BorderManager 2.1 Web Server Accelerator with Support Pack 2D and Proxy Patch 105

Log Mode

Standard


BorderManager

Configured to accelerate Netscape Enterprise Server for NetWare

PROXY.CFG set object cache max hot nodes to 10000 (default 7000)

Added NWAdmin for NT Snap-ins for BorderManager (BMNWNTB.EXE)

Updated MSM.NLM with 12-2-97 version

Updated ETHERTSM with 12-8-97 version

Updated TCPIP.NLM with patch TCPN05

Added FATFREFX.NLM

Added CACHEPER.NLM

Added LOCKPERF.NLM


Cache Volume

NetWare volume spanning four drives, each with a single 400MB partition

8KB volume block size, no suballocation, no compression


Benchmark Configuration

Maximum Requested Load

2150

Maximum File Set Size

1015 MB


Network

Network Controllers

5 Intel EtherExpress PRO/100 Server Adapters

No. of Nets

5, with two clients per net

Type of Nets

100Base-T

Network Speed

100Mb/s

MSL (sec)

30 (Non RFC1122)

Time-Wait (sec)

60 (Non RFC1122)

Notes: Each net connected by a Hewlett-Packard AdvanceStack Switch 800T. All nets at half duplex.

STARTUP.NCF File pmload FATFREFX.NLM pmload CACHEPER.NLM load LOCKPERF.NLM set maximum physical receive packet size = 1514 set maximum packet receive buffers = 2000 set minimum packet receive buffers = 500 set maximum service processes = 100

AUTOEXEC.NCF File set maximum concurrent disk cache writes = 500 set directory cache allocation wait time = 0.1 sec set directory cache buffer nonreferenced delay = 60 min set minimum directory cache buffers = 10000 set maximum service processes = 100 set upgrade low priority threads = off set maximum file locks = 100000

Conclusion

Novell's ability to win SPECweb96 benchmark comparisons against systems with faster processors, even multiprocessor systems, demonstrates the maturity of BorderManager's caching capabilities. This performance and efficiency makes BorderManager's Web server acceleration and proxy caching services a good fit within both Internet and intranet infrastructures when you need to scale up your Web infrastructure.

For complete SPECweb96 submissions and results, see the following Web sites:

Compaq ProLiant 3000 (1639 operations/sec)

http://www.spec.org/osg/web96/results/res98q2/web96-980322-02570.html

HP NetServer LH 3/400 (2131 operations/sec)

http://www.spec.org/osg/web96/results/res98q2/web96-980518-02768.html

Sun Ultra Enterprise 450 (1628 operations/sec)

http://www.spec.org/osg/web96/results/res97q3/web96-970819-02022.html

DEC AlphaServer 1200 5/533 (2045 operations/sec)

http://www.spec.org/osg/web96/results/res98q2/web96-980518-02770.html

* Originally published in Novell AppNotes


Disclaimer

The origin of this information may be internal or external to Novell. While Novell makes all reasonable efforts to verify this information, Novell does not make explicit or implied claims to its validity.

© Copyright Micro Focus or one of its affiliates