Check out the new USENIX Web site. next up previous
Next: Observations Up: Experimental Results Previous: Memory and CPU Overheads

Layer-4 Informed ALRMs

We now explore integrating layer-4 information into ALRMs with speed-sensitive steering, which always selects servers with open flow windows for inbound frames instead of using strict round robin, balancing inbound traffic across the ensemble more effectively. This allows the Anypoint switch to optimistically advertise the sum of the ensemble's flow windows to the peer.

We now compare the inbound throughput of a speed-sensitive steering ALRM versus the TCP proxy. In this experiment there are four servers, server socket sizes (flow windows) are 32KB, and the MTU is 1500 bytes. One server incurs a variable delay between processing frames. Figure 8 shows throughput as this delay increases to 0.75ms. The TCP proxy gates the receive rate of every server to the slowest server. But the Anypoint connection can take advantage of excess capacity at the other servers and maintain high throughput during load imbalances.


  
Figure: With speed-sensitive steering, the ALRM redirects inbound frames to servers with sufficient capacity to handle them. Here one server incurs an increasing CPU delay to process a frame. The TCP proxy gates the receive rate of each server to that of the bottlenecked server.
\begin{figure}
\centerline{\epsfig{file = figs/kens_Speed4.ps, height =1.8in}}
\end{figure}


next up previous
Next: Observations Up: Experimental Results Previous: Memory and CPU Overheads
Kenneth G. Yocum
2003-01-20