Application Performance: The heartbreak of packet loss–cured!

Application Monitor” frequently covers networking at an abstract level, as a component of delivery of applications to end-users, and we’ve just begun to describe the SDN (software-defined networking) trend. Take a moment now to ground those high-level concepts with details of typical low-level packet-loss woes. Andy Gottlieb is in the middle of an examination of existing networking models and their improvement in Next-generation Enterprise WAN (NEW) architecture. He rightly concludes, “Wide Area Network [WAN] design is, for the first time in a long time, strategic.” NEW and allied initiatives including SDN and virtualized networking are also timely, though, for unmistakeably tactical and low-level technical reasons. Gottlieb himself illustrates this with his, “Why does packet loss destroy application performance over the WAN?” Keep in mind that TCP (“transmission control program“, initially, but soon the “transmission control protocol” we recognize today) was originally designed almost forty years ago. As brilliant as its specification was, it was biased toward the span of problems typical in telecommunications of the mid 1970s. The X.25 definition which enabled the early effective WANs was itself only approved a little later, in 1976. In that distant time, TCP effectively encouraged packet loss as a failsafe against excessive congestion on low throughput network circuits. While our technical means now are several generations later, faster, and more reliable, TCP continues to exhibit packet loss in a variety of circumstances. This tendency makes “packet loss … one of the scourges of application performance over the WAN“, in Gottlieb’s accurate words. The result: when packet loss-rates reach levels even as low as 1%, TCP protects the WAN by algorithms that can wreck real-time and interactive application performance.

Non-linear losses; non-linear gains

Gottlieb’s packet-loss WAN problem is just one example of a large collection of non-linearities and interactions in networking. When designing and operating application performance management (APM), it’s important to understand that:

  • response on a WAN might be quite different from the behavior of its constituent LANs;
  • small overloads, in the sense of packet loss rates of a few percent, can result in performance degradations at the application level as large as several hundred percent;
  • depending on the “tuning” of network equipment, TCP loads can interfere with what look like unrelated UDP (user datagram protocol) traffic, and vice-versa;

and even more unusual pathologies. Turn these observations around, though, and they become enthusiasm about “WAN virtualization”, for, much as it already has done with servers and storage, virtualization in networking yields “WANs that have 20 to 100 times the bandwith, with monthly WAN costs reduced by 40% to 80% or even more, and more reliability …” The point for today is that these gains go to the prepared. While the technical pieces–configurable networking equipment, virtualized networks, lively telecomm markets, useful best practices–are available for purchase now, next-generation networking remains enough of a novelty to require real understanding of multi-protocol label switching (MPLS), TCP window scaling, and other implementation specifics. Be clear about your business requirements; make good measurements; and grow or hire the expertise to make the most of networking’s possibilities. In Gottlieb’s inspiring vision, “[o]ne of the most beautiful points is that most of this next-generation network upgrade pays for itself out of the WAN OpEx budget.”