TCP Rate-Halving

Projects
TCP Rate-Halving
NIMI
Autotuning
SACK/FACK
Technology
   Integration

Software
TCP Implementations
TReno
Traceroute
Windowed Ping

Websites
TCP Performance
   Debugging
Performance
   Tuning
TCP Friendly

Related Projects
NLANR
NCNE Engineering
   Services
NCNE GigaPop
PSC
LBNL NRG

Miscellaneous
Staff
Help
Search
Web Feedback

About the PSC Treno Server

Treno (Traceroute RENO) is a network testing tool designed to test network performance under load similar to that of TCP, the most commonly used Transport Protocol in the Internet today.

Treno uses the same technique as traceroute to probe the network. By sending out UDP packets with low TTL, hosts and routers along the path to the final destination will send back ICMP TTL Exceeded messages which have similar characteristics to TCP ACK packets.

Treno also has an ICMP mode, which uses ICMP ECHO Requests instead of low TTL UDP packets. In this mode, you only get information about the final destination; however, the same sized packets are sent in both directions, giving you some information about the return path.

There are several points which people should be aware of when using the Treno server.

• Because Treno simulates the full TCP algorithm including slow start, each hop should run for at least 10 seconds in order to come into equilibrium and get accurate bandwidth measurements. Tests which are too short may give results which are not statistically valid.
• The Treno Server sends back results in real time, using a buffering hack. This results in the server sending back a somewhat large amount of data (~ 5 kBytes per hop), which may be large to people coming in through slow connections. User be warned!
• Some routers in the Internet do not respond to the Treno probes as quickly as they forward packets. This means that the results of the Treno test will not accurately reflect the bandwidth at these hops. The best tests are to high speed Unix hosts which have no difficulty in responding to Treno probe packets as quickly as required.
• The Treno Server is single threaded, so if someone else is running tests you will need to wait until they complete their work.
• The graphs produced by the server use a combination of gnuplot and ppmtogif. If anyone knows a better way to make 2-D datasets into gif or jpeg images (in batch mode), we'd like to hear about it!
• The PSC keeps logging information both for debugging of the server. We may also at some date desire to do statistical analysis on the server logs to characterize Internet performance. If you have a problem with being a part of such an experiment/analysis, please don't use the server.
• As of June 2000, the PSC is connected to the Internet by 4 connections to separate providers. The fastest connection is over a pair of striped DS-3s providing 90 Mb/s of throughput over that path. TReno performance measurement using typical Ethernet packet size of 1500 bytes is meaningful up to approximately 20 Mb/s and is therefore not constrained by PSC's Internet connectivity.

The IETF Benchmarking Methodology Working Group (bmwg) is looking into issues of Internet Provider Performance Metrics (IPPM). The IPPM effort may result in a standard method for using tools such as Treno to provide quantitative performance metrics for Internet Providers. As such efforts develop, we plan to fold them into this server.

This server is being provided by the Networking group at the Pittsburgh Supercomputing Center, in order to encourage Internet users and providers to analyze and improve the performance of their infrastructure. We hope that people will find it a valuable resource. The development of Treno was partially funded by NSF Grant NCR-9415552. If you wish to run Treno at your own site, source code and information are available via anonymous ftp. The file is /pub/networking/tools/treno.shar.