The Entomology of Server Sizing Models

Dr. Neil Gunther

Performance Dynamics Consulting

Abstract

The attractive feature about these laws for performance evaluation stems from the presence of a single parameter in the case of (1-4) or two parameters in the case of (5). These parameters can be determined in a straightforward way using regression on a relatively small set of benchmark measurements.

Compared to higher life-forms, like event-based simulators and performance engineers, these laws appear as ``insects'' (viz, mostly exoskeleton and not much intelligence). On the other hand, like natural insects, these laws can play an important role in the complex biota of performance analysis and capacity sizing. To better understand that role, however, we need to study their anatomy. Unfortunately, the equivalent of Gray's Anatomy has not been written for these performance beasties (although, I did perform some initial dissections in Chapter 14 of my book).

In this talk, I will hold the above collection up to the light and display the following anatomical features:

• None of these entoforms is high enough in the food-chain to be rightly regarded as a ``law'' (in the sense of say, Little's Law).
• Entoform (1) ((Genus Amdahlus) has the correct anatomy of a synchronous queue. It can also be classified as a larval stage of (5).
• Entoform (2) (Bogus Geometrica) has been caught infecting sizing estimates at both Sun and IBM. It is highly poisonous (a Poxian queue) and should not be brought into contact with large-scaly SMPs or clusters. It can be regarded as just a lowly bug!
• Entoform (3) (Scalus Parabolus) is a sub-species of (5)
• Entoform (4) (Species Unknown) was caught infecting sizing estimates at Amdahl Corporation in 1999. It apparently arrived on an asteroid.
• Entoform (5) (Scalus Dromedarius) has the correct anatomy of a single-hump load-dependent queue which includes any possible thrashing effects. It also has a symbiotic relationship with cluster systems.

Without understanding such anatomical distinctions, sizing large-scale multiprocessor servers with these laws-or any others, for that matter-is tantamount to curve-fitting numerology. Predictions (especially those made in the field) may come back to bite you where it hurts most; in the stock ticker!