|
|
|
|
Volera Makes You Cache-Rich Quick
|
 |
|
April 29, 2002
By Sean Doherty
>> continued from previous page
|
|
|
How We Tested ECDN Solutions
|
|
|
To test eCDN solutions at our Syracuse University Real-World Labs®, we set up a private network that emulated an enterprise WAN, then migrated the test devices to that network. We simulated 100 virtual users with Mercury Interactive Corp.'s LoadRunner 7.02 on 500-MHz Celeron-based Dell Computer Corp. computers. These users accessed centralized resources on a corporate backbone over simulated WAN and Internet conditions using HTTP, MMS and RTSP.
We set up a corporate backbone with three servers providing HTTP, MMS and RTSP services. We mirrored Network Computing's Web site (www.networkcomputing.com) onto a dual Pentium III 1,266-MHz PC running Microsoft Windows 2000 server on loan from Alacritech. That server made use of Alacritech's SLIC (Session Layer Interface Card) NICs and 512 MB of RAM. It also served as a Windows Media Server. We made a Sun Microsystems UltraSPARC 10 home to a RealNetworks RealServer 8. Each media server supplied streaming video files encoded at 22-Kbps, 150-Kbps, 450-Kbps and 1,000-Kbps rates and ranging from about 5 MB to 260 MB. Another Windows 2000 server provided DNS, DHCP and NTP services.
We used Shunra Storm 1.1 to create the WAN and Internet conditions. We set up the Shunra Storm to emulate T1 (1,544 Kbps) and T3 (44,736 Kbps) lines. In addition, it emulated Internet conditions over a 100-Mbps network using bidirectional 100-millisecond (standard deviation of 10 percent, packet loss of 1 percent) and 200-millisecond delay (standard deviation of 20 percent, packet loss of 2 percent). The Storm monitored the number of packets that traversed the WAN.
Our test suite included products from Mercury and Keynote Systems. LoadRunner gave benchmarks for client HTTP requests (total hits and hits per second) and verified network conditions with a network monitor. Keynote supplied us Custom Perspective agents for HTTP, RealNetworks and Windows Media. Keynote's information on the quality of streams supplied to clients was not available in time for this review, but we will post the results here. RealNetworks' enterprise-simulation tool, with RealOne Player and Windows Media Player 7.1, gave us baseline streaming-media features and functionality data.
LoadRunner's Virtual User Generator captured a three- to four-minute browser session to our test Web site. We loaded the session into LoadRunner's Controller to play back on 100 virtual clients, using the eCDN cache servers as forward proxy servers. We created a bottleneck on the network, rather than in CPU or RAM, by editing the script to emphasize high bandwidth requirements and cause latency in client applications.
The script ran for more than three minutes against our test site, at www.nwctest.com. Running the script directly at the server under test produced four baseline tests for T1, T3, 100-ms delay and 200-ms delay emulations. We conducted two tests for each emulation and averaged the result. Then we installed each eCDN solution to relieve the congestion.
The script initially requested original content to force cache servers to traverse the WAN/Internet emulation and retrieve original content, return it to cache and serve it to clients. The script then repeated those requests for cache servers to fulfill requests from cache. After about 90 seconds, the script accessed more original content to force the cache servers back across the simulated WAN/Internet to retrieve more content and serve it locally to clients.
|
|
ECDNs: The Next Generation
|
|
|
ECDNs came to market in 2001, but a year later there's still no dominant vendor, according to Gartner. Cache servers are shifting from their point-product pedigree to a network systems approach, to speed content from enterprise data centers to remote locations across slow WAN links and Internet connections of dubious quality.
Rather than simply cache static HTTP and FTP, systems from CacheFlow, F5 Networks, Volera and other vendors support both live and on-demand streaming-media files. Furthermore, with these systems you can pre-position content on cache servers as eCDNs. The evolution from delivering and holding content in cache to creating and managing it has only just begun.
As applications move onto the Web and distance learning and corporate communications flow to desktop media players, cache servers need to support more and more protocols and integrate with back-end support for content management and creation systems. In effect, cache vendors are likely to move to content creation and management to control the entire spectrum of publication, from creation to delivery.
We got a taste of the evolving cache machine. InfoLibria submitted its Content Distribution 5.0 system, with two E-Class MediaMalls all running from 1U appliances under Microsoft Windows 2000 server, with Service Pack 2. The system's Content Operating System version 5 distributes content from origin servers to MediaMalls positioned close to users (see "InfoLibria Rich Media Starter Kit Takes Command of Streaming Media," www.nwc.com/1303/1303sp4.html).
The E-Class MediaMall is an appliance but not a cache server. It doesn't cache frequently used Web content; rather, it hosts Web applications and includes its own application logic and media distribution engines. MediaMall supports Microsoft Windows Media Server, RealNetworks RealServer and Darwin Streaming Server.
InfoLibria's Content Commander and Director come on separate appliances and work together to distribute content to MediaMalls. Content distribution begins by accessing an ASP (Active Server Page) Web site running on the Commander. From the Commander, you create and manage jobs that pre-position content from origin servers on MediaMalls. Jobs identify the content and specify the distribution method to identified MediaMalls on a configurable schedule. From here, the Director redirects client requests for content from origin servers to MediaMalls using global redirection.
Today, InfoLibria integrates with applications like Microsoft Producer to publish streaming media to MediaMalls close to end users. Tomorrow, it may be working on your final frontier.
Content Commander 5.0, MediaMall E-Class 5.0, Director 5.0, InfoLibria, (781) 392-2200.; www.infolibria.com
|
|
KeyNote Results
|
|
|
Notes
To judge the quality of streams and HTTP performance of our Editor's Choice Content Delivery Solution, we compared the output of a Volera Excelerator version 2.1 with a Solaris 8 server running RealNetworks RealSystem Server 8 and an MS-Windows 2000 server running Windows Media Services 4.1 and IIS (Internet Information Services). Using Keynote Systems Perspective Agents, we found little difference in the quality of streams and no difference in Web performance on a 100-Mbps LAN.
Keynote Systems shipped us two Streaming Perspective Agents to monitor Windows Media Services and RealNetworks media and a Web Site Perspective Agent to monitor Web-site performance. The streaming agents measured the quality of streams from origin media servers and from a Volera Excelerator cache server. The Web-site agent measured the response times for Web objects from an origin Web server and from the cache server.
Keynote Agents requested Web objects and 22-, 150-, 450- and 1000-Kilobits-per-second encoded streaming media files directly from origin servers at 30-minute intervals for two days. Then, we inserted an Alteon 184 switch between the Keynote Agents and the origin servers.
We configured the switch to filter off original content requests from Keynote Agents and send them to an Excelerator cache server in transparent mode. All Web objects and streaming media files were prepositioned and ready for action on the Excelerator. After another two days of data gathering at 30-minute intervals, the origin servers and the Excelerator played to a draw.
We found no difference in the response times of our origin Web server and the Excelerator. Both averaged .03 seconds response time over a 100-Mbps LAN.
The Excelerator actually performed a little better in delivering Windows Media to agents than the origin server, but it did not fair as well against the RealSystem Server. But the small differences may not be visible to end-users. Both origin servers and the Excelerator ran neck and neck in the frames-per-second category.
Performance Testing: Keynote Systems, Web Site Perspective Agent and Streaming Media Perspective Agents for Real and MS-Windows media
Key
Stream Quality: a 0-10 rating for overall experience of playing a stream, with 10 being the highest quality. This rating is derived from a combination of Startup Score, Audio Score, Video Score and the encoding factors.
Video Score: a 0-10 rating of the video portion of the stream. It takes into account encoding factors such as encoded bit rate and frame rate as well as delivery and rendering factors such as bandwidth delivered and packet information.
Rendering Score: a 0-1 qualitative rating of how well the streaming audio/video was rendered at the receiving end, regardless of encoding factors. Therefore, two streams of different sizes that are delivered and rendered perfectly will receive the same rendering score of 1.
Startup Time: This is a combination of three things: (1) initial connection time (includes DNS and time to first byte); (2) redirection time; (3) initial buffering time.
Frames per Second: shows the average frame rate received by the Keynote agent.
Bandwidth: displays the average audio and video bandwidths delivered to the Keynote agent during playback.
|
|
|
|
|
RSS
