03:44 pm, 18 May 06
latency and predictive interfaces
It's well known that computers are generally "fast enough", at least in terms of throughput, but what I care about more as a user is latency. Latency is hard to eliminate: if clicking a button really does require a bunch of computation (like sorting a list), then you can optimize the algorithm to a certain point and then you're stuck.
Two thoughts on the subject:
Two thoughts on the subject:
- Does anybody measure interface latency on native apps? For Ajax apps people put a lot of effort into measuring latency, because it's typically much higher there. But native app latency matters too, and it feels like it's been increasing. See Glen's notes on trying to switch to OS X -- as I commented there, it would seem to me there's probably a window of time where your brain interprets "key press here + visual change there = CAUSALITY" at a very low-level way.
- How could you bring latency even lower? Start the computation before the user clicks! If you precompute the result of the most likely actions, then invoking that action can happen nearly instantaneously. This really depends on how much IO is involved, of course, and maybe most latent actions involve IO. But when I'm grumpy about latency, it's for stuff where the IO overhead isn't significant: like running "ls" in a directory right after I "cd" into it, or clicking "compose new mail" in an email client.
most likely actions - i like the idea that eccentric users are penalized for interacting with apps in an uncommon way.
That these boundries are so discrete is pretty fascinating. Only tangentially related: something I learned from using satphones and VoIP-over-VSAT is that there is a point right around when audio rtt reaches 500ms that people instantly start using it like a half-duplex walky-talky instead of a phone, and the format (and duration and frustration level) of the discussion totally changes. (The amount they'll willingly pay for the call also drastically decreases.) You can of course decrease the size of the jitter buffers, but there's a similar boundry for that failure condition, as most mobile phone users are well aware of.
People like seeing their CPU usage low, even though it really means they're wasting resources they could be using.
My favorite latency annoyance is Thunderbird's "compose new mail" latency. The window pops up pretty quickly, and you can start typing right away. But there's a significant delay before the cursor moves from the body textarea to the To: field.
Honestly, it makes me question the viability of XSL as a toolkit outside of Mozilla-et-al. There's no way you'd leave that bug in Thunderbird if you found it, so it almost certainly is a problem with something lower.
re: 1
One method is to use the model human processor. IIRC, it's something like 70ms +- 30ms between click and expecting something to happen. Likewise, video only needs something like 15-20fps to appear continuous.