Balancing server loads globally.
Internet World, June 7, 1999 v5 i21 p27
Author
Phifer, Lisa
Summary
Distributing data worldwide for the Web typically relies on load balancing thatrrangements that involve
regional work centers. Content providers distribute their data repositories nationally and internationally.
Multinational organizations need to provide fast, consistent response times, without having to mirror entire
Web sites. Online configurations need to be transparent for users, and no amount of downtime is regarded as
acceptable.
Full Text
Bringing data closer to users means faster access times, less downtime, and greater manageability
DATA, LIKE POLITICS, is local, since it has to end up somewhere. And while the Internet's distributed nature
means that the physical location is hidden from end users, all things being equal, it still takes more time to
bring data that's farther away than it does to get at local data.
But distributing data worldwide for the World Wide Web is a way to provide faster service to targeted
populations, localization, and more robust service.
Server load balancing is nearly as old as the Web itself, with roots in the round-robin DNS once used to
distribute HTTP requests evenly across a pool of servers. By 1997, the first generation of load balancing
products began to emerge, offering algorithms to better utilize Web servers at a single location. Products like
Cisco's Local Director, F5 Network's BIG/ip, and Radware's Web Server Director were deployed in front of
server farms, providing one "virtual IP" address for the entire site.
Packets arriving at this virtual address were forwarded to "best" destinations using such metrics as server
availability, response time, and user-defined weights. These solutions allowed server farms to scale
transparently and to become resilient to single-server outages.
But load balancing from a single server farm still leaves a site at the mercy of every connection leading to it.
It's like building the perfect store served by a single road.
This is why real-time, transaction-intensive sites such as E*Trade now involve more than one regional work
center. Content providers like USA Today distribute data repositories nationally and internationally.
Companies that operate globally want to provide consistent response time to visitors anywhere from Bangkok
to Boston, without having to mirror entire Web sites.
"You've got to make it transparent for users to get to the closest content--a single site is no longer an
acceptable way of doing business, and the notion of acceptable downtime is going away," said John Stewart,
director of systems engineering and security at Digital Island, a high-speed overlay network service provider
with four international data centers.
Nearly every local load balancing product now sports a global counterpart or add-on. But considerable
diversity exists in this rapidly emerging market.
The granddaddy in this arena is Cisco Systems Inc.'s Distributed Director, which turns Cisco 2500 and 4700
routers into global load balancers. Companies like Digital Island deploy a pair of Distributed Directors for
redundancy, supported by a Local Director at every data center.
Stewart prefers Cisco's approach "because it understands the network layer, and is engineered from the
bottom up, instead of top-down." Distributed Director calculates network proximity by querying routers for
BGP and IGP route info, then combines round-trip latency, server up/down status, and administrative input to
select the "best" server. Two modes of operation can be used: HTTP redirection, by returning a "302
Temporarily Moved" response, or redirection of any application using DNS resolution.
Distributed Director is a relatively mature, stable product that leverages finely tuned network layer distance
metrics. But it requires Cisco routers at every site, plus BGP peering, and it doesn't take into account current
server workload.
Radware Inc.'s Web Server Director for Network Proximity takes the appliance approach, providing a
dedicated box for load balancing. WSD-NP, like Cisco's Distributed Director, also supports HTJ'P and DNS
redirection, but adds a third method, called Triangulation, whereby one box redirects traffic to another. The
second WSD acts as a proxy, returning responses directly to the requesting client. Radware designed
Triangulation as a high-throughput any-protocol alternative, because DNS redirection works well only if the
DNS server is geographically close to the client-a particularly bad assumption for road warriors.
"Our job as a vendor is to provide flexibility--no two clients think alike," said Hooman Beheshti, Radware's
chief technical officer. "We allow customers to choose the role of each WSD, redirection method, algorithm
metrics, and failover configuration." Any WSD Pro can be upgraded to an NP; NPs can perform both local and
geographic load balancing within a single box, and redundant NPs can share the balancing workload.
Radware and F5 Networks Inc. both offer local balancers that measure server workload using such metrics as
the number of open connections, fastest response time, number of successful requests, and packet
throughput. But F5's BIG/ip adds content awareness to the mix. For example, it can redirect around "404
Object Not Found" messages that might otherwise be interpreted as fast, successful responses. Extended
Content and Application Verification tools allow entire transactions to be emulated, with test results factored
into an "Internet quality of service" algorithm.
ENSO, a BIG/ip user that distributes audio clips to customers like Tower Records, deployed F5's 3DNS to
prevent unacceptably high packet loss over WAN links. F5 argues that dedicated, specialized hardware is
necessary to sustain reliably high throughput under stress. "Ninety-five percent of our sales involve
high-availability configurations; this underscores our belief that these solutions must not become a single
point of failure," said F5 director of product management Dan Matte.
Coyote Point Systems Inc.'s chief engineer, Bill Kish, agreed. "Disaster recovery is the primary issue pushing
geographic load balancing," he said. Coyote Point's Equalizer, he said, has attracted an e-commerce customer
base that will probably want Envoy, an inexpensive geographic balancing add-on that starts at just $2,500
per site.
Coyote Point customer IMDb (Internet Movie Database) has been using Envoy to serve images from the nearest
U.S. or U.K. site for six months. "Our sites had no way of covering for each other until Envoy came along," said
Jake Dias, IMDb's systems manager. "We are now able to offer quick service to all users, wherever they are.
Any site can go down and nobody will even notice."
Server load balancing modules are also available for "wire speed" switches like Alteon's ACEdirector, Foundry
Networks' ServelIron, and ArrowPoint's Content Smart Web Switch, These products are designed to move LAN
traffic via high-density Fast and Gigabit Ethernet ports and ASIC-based Layer 2/Layer 3 switching. Layer 4
redirection software has been added to support various applications, from load balancing to Web caching.
For example, Alteon's WebOS global server load balancing option allows ACEdirector switches to redirect
traffic based on server health, proximity, and response time. Foundry's Internet IronWare option supports
global server load balancing as well.
Where do these switching products fit? Mike Long, vice president of marketing and technology at Radware,
predicts that switches will eventually subsume the local load balancing market, while special-purpose
balancing products will reign in the distributed market and in LANs where intelligence takes precedence over
speed.
What other innovations are we likely to see in the next generation of load balancing products? Content
awareness will continue to grow, spurring products that understand how enterprise applications behave
end-to-end. An example of this trend can be seen in Resonate's Central Dispatch, a load balancing product
that evaluates the health of not only the target Web server, but also the back-end server required to satisfy an
incoming HTTP request.
Sri Chaganty, vice president of engineering at HolonTech, predicts that switch vendors will consolidate
value-added functions, such as quality of service rate-shaping, bandwidth management, and other
access-layer services, while embedding basic load balancing in ASICs. Some switches may become more
tightly coupled with the server farms they front-end, embracing new operating-system load balancing
features such as Microsoft's Clustering Services.
Today's two-tiered products rely on proprietary communication between global and local balancers to
determine proximity, network, server farm, and host performance. Best-of-breed multivendor combos
pairing high-speed LAN switches with intelligent software load balancers will require industry cooperation
and partnership. But this awaits a number of new ideas the players are still investigating.
Coyote Point's Kish thinks there's another step that will eventually become critical: to proactively push content
where it's needed, before it's requested, Radware's Long hypothesized that "reverse proxy caches" might bring
content closer to users by augmenting or replacing mirrored sites with cached content. To exploit these
resources, load balancers must become smart enough to differentiate between cachable and non-cachable
content requests.
If enterprises adopt global load balancing to provide bulletproof 24-by-7 network presence, the load
balancers must themselves be rocksolid and secure. As this market matures, high-availability configurations
deployed in redundant pairs may become the norm, and greater emphasis will be placed on security. The
more sophisticated customers will demand management tools that help them evaluate traffic, predict growth,
and tune policies for optimal performance, while customers at the lower end of the market will demand
self-tuning turnkey "appliances" that can be dropped into a network with minimal fuss. |
