Submission + - Plummeting BDP on High Speed Lines
Lawrence_Bird writes: In trying to track down a tcp problem today I've stumbled upon what seems to be a nasty little issue — incredibly low ping times!
I have a 25Mbps line and in checking RTT to places I frequent, I found that the times are tanking. Not that long ago, a fast link was like 60ms with most around 90-120ms, twice that overseas. Now many are coming in the 45-80ms range. But quite a few others, like youtube (6ms) google (7.5ms), apple (15ms) cbs.com (20ms) slashdot (30ms) , are a lot shorter.
This presents a very difficult problem. How to best tune RWIN? The rule of thumb had been to make rwin just greater than your typical BDP and most information on the topic is written at a time when high speed lines were starting to appear but latency was still high (>100ms). But now we have a situation where yes, rated speeds are high, but latency has dived to next to nothing for major internet sites but may still be high(er) elsewhere creating not a 2 or 3 to 1 range of RTTs but instead as high as 25 or 30x!
To optimize for google would require setting it to under 24,000 bytes. In fact, any average ping time below 48ms will (I think) saturate if RWIN is over 65,535 (the point at which scaling would take over). Yet there are still a goodly number of places that are over 80ms (say stanford.edu) and of course europe and asia can be 200ms.
So I am curious what others are doing to tune RWIN to provide the necessary throughput on the longer latency runs against these now incredibly short pings
I have a 25Mbps line and in checking RTT to places I frequent, I found that the times are tanking. Not that long ago, a fast link was like 60ms with most around 90-120ms, twice that overseas. Now many are coming in the 45-80ms range. But quite a few others, like youtube (6ms) google (7.5ms), apple (15ms) cbs.com (20ms) slashdot (30ms) , are a lot shorter.
This presents a very difficult problem. How to best tune RWIN? The rule of thumb had been to make rwin just greater than your typical BDP and most information on the topic is written at a time when high speed lines were starting to appear but latency was still high (>100ms). But now we have a situation where yes, rated speeds are high, but latency has dived to next to nothing for major internet sites but may still be high(er) elsewhere creating not a 2 or 3 to 1 range of RTTs but instead as high as 25 or 30x!
To optimize for google would require setting it to under 24,000 bytes. In fact, any average ping time below 48ms will (I think) saturate if RWIN is over 65,535 (the point at which scaling would take over). Yet there are still a goodly number of places that are over 80ms (say stanford.edu) and of course europe and asia can be 200ms.
So I am curious what others are doing to tune RWIN to provide the necessary throughput on the longer latency runs against these now incredibly short pings