Presentation on theme: "GSFC to Alaska Performance Results Tino Sciuto Swales Aerospace ESDIS Network Prototype Lab. NASA GSFC Greenbelt, MD."— Presentation transcript:
GSFC to Alaska Performance Results Tino Sciuto Swales Aerospace ESDIS Network Prototype Lab. NASA GSFC Greenbelt, MD
Motivation Characterize network service parameters for future purchase of network services to Alaska Baseline TCP Throughput between NASAs Goddard Space Flight Center in Greenbelt, MD and Alaska Identify network as well as end-to-end performance issues Compare SGIs with Linuxs TCP stack performance –Measure performance with different parameters SGI specific like tcp_gofast Global TCP parameters like, window sizes and buffer length
University Of Alaska PNW GigaPoP MAX GigaPoP GSFC Research LAN Abilene
Test Info Sender Hosts –SGI OS IRIX 6.5 OC-3 ATM NIC with MTU 9180 –End to end MSS was 1460 SACK enabled –Linux Red Hat 7.2 GE NIC with MTU 4470 –End to end MSS was 1460 SACK enabled Receiver host: no specific info available Test tool nuttcp –For IRIX : nuttcp version 3.1.9 –For Linux : nuttcp version 3.1.6 Test Methodology –At least two minutes test runs every hour. Runs repeated on as needed basis Wanted to reduce impact on Alaska connection (i.e. would like to continue testing ;-) –Monitored MRTG statistics on routers –Recorded routes, ping avg. and TCP throughput –Varied window and buffer length parameters to identify effects on performance –Tabulated results Troubleshooting –Started using web100 tool in the last week to study TCP behavior –Many parameters to track (SACK, Retransmittions ….)
Linux Test Results Excessive Window Deteriorates Performance Link Capacity ??
Conclusions Operation of SGI TCP stack not fully understood –Larger nuttcp buffer length results in better performance …? –Without tcp_gofast > RTT even increasing window sizes didnt help much the SGI host tcp_gofast essential to obtaining maximum throughput –Web100 for IRIX not available Linux TCP stack produced expected results –Alaska to Seattle link might be OC-3 Results seem to indicate a packet loss between E-5 and E-6 when comparing the results with the expected results using the formula Bandwidth = 0.7 * Max Segment Size Round Trip Time * sqrt[packet loss] Mathis, et. al