BOSTON (05/08/2000) - Eliminating Web delay is a huge corporate mantra these days. We recently tested Phobos Inc.'s new ipXpress - a compact, four-port load-balancing switch - to see whether the company's claims that it could blow away Web delay are warranted.
We found that ipXpress - which began shipping last month - effectively handles moderate-to-high Web traffic loads, those in the 1,000 hit/sec range, and successfully supports a number of different load-balancing methods that can help keep your server cluster humming and your e-customers happy.
IpXpress supports six different load-balancing algorithms - ranging from equal "round-robin" allocation to more complex "weighted, least-connection" load distribution. The range of supported load-balancing options is fairly rich, especially given the unit's low price of $3,495. Other load balancers on the market, with equivalent functionality, cost $10,000 or more.
This unit is a low-end Web load balancer designed for sites with three or four servers. Increasingly, multiple servers comprising a single Web site are collocated in rack space rented from a service provider. IpXpress supports four 10/100Mbit/sec Ethernet ports. These ports can be a mix of inbound and outbound traffic paths, but one port is automatically used for inbound traffic - the connection out to the IP cloud. Given its low price, ipXpress seems well suited for e-commerce and dot-com sites, in which support for three or four Web servers is sufficient.
We note that ipXpress can support more than just three servers, although no more than three with dedicated 100M bit/sec connections. We accomplished this added scalability in our testing by first connecting the load-balancer unit to a switch, then connecting six Web servers to the switch. We found this worked well without impacting the overall performance of ipXpress - at least not for the moderate volumes of Web traffic we applied. In such a topology, though, all the servers share the same 100M bit/sec of bandwidth back to the IP cloud. In a very high-volume Web site, a shared 100M bit/sec connection may be insufficient and could become a point of congestion.
We conducted tests to see if there were any differences in server processing or user response time for a directly accessed Web server, compared to one connected through ipXpress. Here we were looking at whether adding ipXpress into the network had any bearing on traffic; it didn't. We saw no discernible difference in performance or latency between the two environments.
As far as load, we determined that ipXpress could readily process 795 Web transaction/sec, which is the maximum load we could apply in our test environment. A transaction consists of a completed Web request and download.
The load-generating software we used delivered Web files (pages) in varying sizes and numbers, ranging from 5K bytes to 5M bytes per download - most were between 10K and 20K bytes. This amount of traffic equated to a Web-download traffic volume that averaged about 11M bit/sec.
We believe the Web traffic volume we applied represents a moderate-to-heavy load for three Wintel-based Web servers. Phobos' documentation implies that its load balancer should readily handle at least 1,000 Web transaction/sec, and based on our testing, we have no reason to question that claim.
In all, we tested how well four of the six supported load-balancing algorithms worked. The first was a test of round robin, in which Web load is equally distributed among the servers. We tested this algorithm with up to 400 Web browser clients and concluded that ipXpress performs as expected, distributing the load equally among six Web servers.
Next we checked "weighted least-connection" load balancing, in which ipXpress balances load based on the number of requests coming in to the servers. It also factors in weighting, where the higher the number, the higher the priority a server has for getting traffic. We set up a traffic ratio of 4 to 1, with two servers weighted at one, and a third server weighted at four. We concluded that ipXpress performs as expected: The server weighted at four received proportionately more traffic than the others (this algorithm also takes into account the number of current connections).
We also tested "least connections" and "fastest response time" load balancing.
Both performed as expected. With the fastest response time algorithm, ipXpress continuously monitors the response times of all connected servers and sends connections to the machine with the fastest response at any given moment. Using the least connections algorithm, ipXpress directs requests to the server with the least number of open connections.
Other balancing techniques
IpXpress also supports two other load-balancing techniques: "adaptive" (in which connections are directed to the servers with the least number of open and pending connections); and "fixed" (in which incoming requests are directed according to source IP address). We could not test these latter two very complex load-balancing algorithms.
Besides load balancing, ipXpress offers other administrative and fault-tolerant features. For example, it constantly monitors the status of connected servers.
It can also be configured in failover mode, in which traffic is routed around a server that goes down. The product also supports redundancy for itself, where one ipXpress box takes over for another in the event of failure. In addition, port redirection to specific ports for FTP or Secure Sockets Layer (SSL) transfers is also supported. SSL is CPU-intensive, so if it's only a small portion of the traffic, it's a good idea to push it off to another port. FTP is bandwidth intensive and, again, if it's only a small portion of the traffic, it's a good idea to redirect it to another port as well.
IpXpress is small, measuring only 8 by 9 by 2 inches and is clearly designed for desktop use. The midnight blue enclosure houses the power supply and processor. LEDs monitor the product's power, link status and overall activity.
We note that the unit would be difficult to rack mount because of its nonstandard size and shape.
We configured ipXpress using Phobos' configuration tool, which supports a command-line interface for installing and configuring the product. The Phobos ipXpress management software resides on a separate PC that is located on the same LAN segment as ipXpress. This software immediately detected the appliance.
We easily walked through the steps outlined in the Phobos documentation manual to add an IP address for ipXpress, and a virtual IP address (what the outside world sees as the Web address). We then added the servers that shared the load of that Web address and configured all other features without any problems.
Documentation, provided in a PDF, is well designed and helps complete initial set-up and configuration quickly. One criticism is that the documentation doesn't specify as clearly as it should that overwriting a command will change it. We found out the hard way.
If you're looking to improve server performance and reduce Web site access delays by adding a load balancer but can't afford a high-priced product, check out ipXpress. It provides good functionality, and with its array of load-balancing techniques, failover and redirection operations and redundancy, it is fully featured.
Yocom is senior editor and Hommer is project manager at Mier Communications, a network consultancy and product test center in Princeton Junction, New Jersey.
They can be reached at betsy@ mier.com or mhommer@ mier.com, respectively.