|X11PERF(1)||General Commands Manual||X11PERF(1)|
Many graphics benchmarks assume that the graphics device is used to display the output of a single fancy graphics application, and that the user gets his work done on some other device, like a terminal. Such benchmarks usually measure drawing speed for lines, polygons, text, etc.
Since workstations are not used as standalone graphics engines, but as super-terminals, x11perf measures window management performance as well as traditional graphics performance. x11perf includes benchmarks for the time it takes to create and map windows (as when you start up an application); to map a pre-existing set of windows onto the screen (as when you deiconify an application or pop up a menu); and to rearrange windows (as when you slosh windows to and fro trying to find the one you want).
x11perf also measures graphics performance for operations not normally used in standalone graphics displays, but are nonetheless used frequently by X applications. Such operations include CopyPlane (used to map bitmaps into pixels), scrolling (used in text windows), and various stipples and tiles (used for CAD and color half-toning, respectively).
x11perf should be used to analyze particular strengths and weaknesses of servers, and is most useful to a server writer who wants to analyze and improve a server. x11perf is meant to comprehensively exercise just about every X11 operation you can perform; it does not purport to be a representative sample of the operations that X11 applications actually use. While it can be used as a benchmark, it was written and is intended as a performance testing tool.
As such, x11perf DOES NOT whittle down measurements to a single ``HeXStones'' or ``MeXops'' number. We consider such numbers to be uninformative at best and misleading at worst. Some servers which are very fast for certain applications can be very slow for others. No single number or small set of numbers are sufficient to characterize how an X implementation will perform on all applications. However, by knowledge of your favorite application, you may be able to use the numbers x11perf reports to predict its performance on a given X implementation.
That said, you might also want to look at x11perfcomp(1), a program to compare the outputs of different x11perf runs. You provide a list of files containing results from x11perf, and it lays them out in a nice tabular format.
For repeatable results, x11perf should be run using a local connection on a freshly-started server. The default configuration runs each test 5 times, in order to see if each trial takes approximately the same amount of time. Strange glitches should be examined; if non-repeatable one might chalk them up to daemons and network traffic. Each trial is run for 5 seconds, in order to reduce random time differences. The number of objects processed per second is displayed to 3 significant digits, but you'll be lucky on most UNIX system if the numbers are actually consistent to 2 digits. x11perf moves the cursor out of the test window; you should be careful not to bump the mouse and move it back into the window. (A prize to people who correctly explain why!!).
Before running a test, x11perf determines what the round trip time to the server is, and factors this out of the final timing reported. It ensures that the server has actually performed the work requested by fetching a pixel back from the test window, which means that servers talking to graphics accelerators can't claim that they are done, while in the meantime the accelerator is painting madly.
By default x11perf automatically calibrates the number of repetitions of each test, so that each should take approximately the same length of time to run across servers of widely differing speeds. However, since each test must be run to completion at least once, some slow servers may take a very long time, particularly on the window moving and resizing tests, and on the arc drawing tests.
All timing reports are for the smallest object involved. For example, the line tests use a PolyLine request to paint several lines at once, but report how many lines per second the server can paint, not how many PolyLine requests per second. Text tests paint a line of characters, but report on the number of characters per second. Some window tests map, unmap, or move a single parent window, but report on how many children windows per second the server can map, unmap, or move.
The current program is mostly the responsibility of Joel McCormack. It is based upon the x11perf developed by Phil Karlton, Susan Angebranndt, Chris Kent, Mary Walker, and Todd Newman, who wanted to assess performance differences between various servers. Several tests were added in order to write and tune the PMAX (DECStation 3100) servers. For a general release to the world, x11perf was rewritten to ease making comparisons between widely varying machines, to cover most important (and unimportant) X functionality, and to exercise graphics operations in as many different orientations and alignments as possible.
|x11perf 1.6.0||X Version 11|