6

Basically, I have an application that does searches and requests results from a backend via long polling. It connects to the backend which collects results for 500 ms then sends them back to the client (I'm simplifying things here a bit of course). As we've optimized this, we've noticed a growing divide between users who search at night vs users that search during the day. Namely, users who search at night are much slower than users that search during the day due to the fact that these users are mostly international and have higher latency.

What are good high-level approaches to speeding things up for these users?

Jason Baker
  • 9,625
  • 8
  • 44
  • 67

2 Answers2

6

You need to analyze your code looking for "turnarounds". A turnaround is a case where forward progress is blocked until data can be received from the other side. Each turnaround adds a penalty to performance that scales with the latency. So if you have 12 turnarounds, 100ms more latency means 1.2 seconds of wait. If you can drop the turnarounds to 4, 100ms additional latency only means an extra .4 seconds of wait.

Some turnarounds are explicit. If you send a request and wait for a reply, that's a turnaround. If you need to see that reply to decide what the next request will be, that's a turnaround.

Some turnarounds are implicit. If you open a TCP connection and wait to determine if the opening succeeded, that's a turnaround. If you close a TCP connection and wait for the close to complete with no errors, that's a turnaround.

Finding and measuring turnarounds is a bit of a black art. But one trick you can use is a latency simulator (Linux can do this with netem). If you can add logging with timestamping or otherwise tell what's going on, you can add a 4 second latency simulator (a device that adds 4 seconds of delay to every packet passing through it). Each 4 seconds you spend waiting for something is turnaround. You can count them, and you can also see where in your operating process they fall.

Code that's never been optimized for turnarounds frequently has a large number of needless turnarounds. They're typically invisible in testing environments because those environments typically have the client and server on the same LAN or, at worst, very close on a fast WAN.

There are four basic ways to eliminate a turnaround. One is pipelining -- this basically just means you don't wait for a reply if you don't have to. You can come back and check results later. Another is consolidation -- if you currently need two requests to do some higher-level operation and you reduce it to one request, that's a turnaround gone. The third way is concurrency -- if you can do more than one thing at a time, that's a turnaround gone. The last is prefetching -- if you know a web page element you're about to load is going to need to load the contents of another URL, you can start retrieving that URL while you're loading the web page that needs it.

David Schwartz
  • 4,676
  • 22
  • 26
0

Reduce the amount of data you're sending over.

Some ideas that might help:

  • Pay close attention to white space in your html
  • Enable compression in your web-server for static pages
  • Split up your responses such that they can be sent gradually to the user
  • Implement an animated waiting/loading screen, which is quick to transfer
blueberryfields
  • 13,200
  • 8
  • 51
  • 87
  • 1
    Which of those things do you think helps with high latency? These all seem to be things aimed at optimizing for low bandwidth. – David Schwartz Sep 22 '11 at 05:35
  • In many cases splitting your responses would have a negative affect rather than a positive one based on the size and the overhead of the responses. – AlexC Sep 22 '11 at 08:26