News for hackers. Stuff that matters.
Re:Introducing Brotli: a new compression algorithm for the i... by cbr on Tuesday September 22, 2015 @09:49AM (#10258493)
The metric divides compression ratio by log compression time. (Ignore for now the normalization to a standard compressor.) This is:

    r / log(T)
That doesn't seem like a good metric, because it overvalues changes in T. For example, say we currently can manage 10% compression (ratio = 100/90 = 1.11) and it takes us 16ms. That's

    r / logT = 1.11 / log(16) = 0.2775
Now we have two proposals. One brings us to from 10% compression to 55% compression (ratio = 100/45 = 2.22) while the other one drops compression time to 4ms:

    r / logT = 2.22 / log(16) = 0.555
    r / logT = 1.11 / log( 4) = 0.555
But improving compression by 5x matters more than improving speed by 4x. Take this to the extreme: a compressor that exits immediately leaving its input unchanged has the best possible score here, despite being useless.

There are no comments on this comment yet.