Check out the new USENIX Web site.
...Banga
This work supported in part by National Science Foundation Grant CCR-9503098
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...call
We use the term receive system call to refer to any of the five system calls available to read data from a socket. The term send system call is used analogously to refer to system calls that write data to a socket.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...packets
Similar problems can arise under load from TCP connection establishment request packets.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...attacks
Often, a denial-of-service attack is used as part of a more elaborate security attack.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...queue
The present discussion assumes that the network interface has an embedded CPU that can be programmed to perform this task. Section 3.2 discusses how LRP can be implemented with an uncooperative NI.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...process
For a shared or multicast socket, this is the highest of the participating processes' priorities.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...interrupt
With soft demux, a host interrupt always occurs upon packet arrival.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...further
With soft demux, the throughput diminishes slightly as the offered load increases, due to the demultiplexing overhead.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...application
In UNIX, more than one process can wait to read from a socket. In this case, the process with the highest priority performs the protocol processing.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...forwarding
QoS attributes or IPv6 flows could be used in an LRP based IP gateway to provide more fine-grained resource control.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...network
This is to ensure that the UNIX scheduler does not consider these server processes I/O-bound, which would tend to give them higher scheduling priority.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...throughput
Note that a (slow) T1 link is capable of carrying almost 5000 SYN packets per second. With the emerging faster network links, routers, and a sufficiently large user community, a server could easily be subjected to such rates.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

Peter Druschel
Mon Sep 16 18:13:25 CDT 1996