• A high precision timer which provides finer granularity timing of the probe packets. Specifically, where the usual timer resolution of a system clock is tens (or more) milliseconds, these tools are based on an SGI memory-mapped device having a resolution of 40 nanoseconds.
• The ability to change the priority of the process to facilitate accurate timing. Basically, the priority of the process making the measurements is elevated to prevent it from being context-switched out while measuring.

• clink uses the even-odd technique described in the SIGCOMM paper to generate interval estimates for bandwidth.
• when clink encounters a routing instability, it collects data for ALL the paths it encounters, until one of the paths generates enough data to yield an estimate.

• TCP and UDP via BSD Sockets
• DLPI
• Unix Domain Sockets
• Fore ATM API
• HP HiPPI Link Level Access

SProbe is a tool for measuring bottleneck bandwidth in an uncooperative environment, i.e., one in which measurement software is deployed only on the local measurement host. SProbe uses the packet pair technique and exploits properties of the TCP protocol in a manner inspired by Savage's Sting tool. SProbe takes no more than three round-trip times (RTT) to produce a single estimate.

SProbe can measure bottleneck bandwidths in both directions of a network path. To measure bottleneck bandwidth to a remote host, using its default settings, SProbe sends six TCP SYN packets to an inactive port on the remote host and receives six TCP RST packets. Several heuristic tests check for abnormal packet arrival times to detect cross traffic -- foreign traffic that can alter the accuracy of an estimate. For a single measurement, SProbe generates 3160 bytes and receives 240 bytes.

To measure bandwidth from a remote host, SProbe relies on application level protocols. Currently, SProbe contains protocol modules for measuring bottleneck bandwidth from Web servers and Gnutella peers. For Web servers, SProbe generates an HTTP get request with a large TCP maximum segment size (MSS). SProbe waits for the Web server to send two large, back-to-back packets and measures their interarrival times producing a bottleneck bandwidth estimate.

SProbe is distributed with source code and installation instructions. It was tested on Linux platforms with kernel versions 2.2.x and higher and FreeBSD 3.4 and 4.2 platforms. SProbe uses the up-to-date libpcap library (ver. 0.6.2), included in SProbe's distribution. SProbe also uses a software firewall, ipfw, supported in typical installations of RedHat 7.x and FreeBSD 4.2. Finally, SProbe makes use of raw sockets, and therefore needs to execute under root privileges.

Much of the code SProbe builds on is borrowed from Sting.

• Optional ttl reporting.
• Optional use of the loose source routing facility, to show the route between arbitrary destinations.
• Enhanced portability, to run on a variety of platforms.
• Improved timeout handling during icmp packet catching.
• Option to probe all addresses of multi-homed destinations.
• Option to disable fragmentation and perform MTU discovery.
• Recognize various new ICMP packet types.
• Round-trip time reporting in fractional milliseconds.
• Option to display the Autonomous System number for each hop.
• Option to display the network name for each hop.
• Configurable default options via environment variables.
• Optional setting of initial ttl to skip first hops.
• Optional min/avg/max rtt statistics summary for each hop.
• Include standard deviation in rtt statistics summary.
• Cache nameserver lookups to minimize DNS queries.