140 likes | 150 Views
This workshop explores the behavior of bbftp in network testing, specifically focusing on the number of parallel transfer streams, packet compression, buffer size, and input/output window size. The effects of different parameters on transfer rate and the impact on other users are analyzed.
E N D
Network Testing Using bbftp Tomasz Wlodek University of the Great State of Texas at Arlington
The behavior of bbftp is determined by several parameters: • Number of parallel transfer streams (default: 1) • “On the flight” packet compression during transfer • Buffer size • Input/output window size Let us change the parameters and see how it affects transfer rate! Maybe the default settings are not optimal? Tomasz Wlodek DØRACE Workshop
Number of parallel transfer streams • bbftp can transfer data in multiple parallel streams. • Transfer a file from hep.uta.edu to d0test.fnal.gov. Each time I change number of parallel streams. • Transfers with and without compressing data • Repeat them several times, at day and night and average the results Tomasz Wlodek DØRACE Workshop
Tansfer rate for various number of paralel streams Tomasz Wlodek DØRACE Workshop
Transfer rate with data compression No Clear Effect Tomasz Wlodek DØRACE Workshop
Ok, transferring in parallel increased transfer speedBut how about other users? Do they feel the effect? • Let user A transfer a file from UTA to FNAL • We will measure the transfer rate with 1 and 10 streams. • Then we let user B open simultaneously transfer with 10 streams. • What will be the effect on user A – with 1 and 10 streams? Tomasz Wlodek DØRACE Workshop
90% If user A runs 1 stream and user B opens 10 additional streams then transfer rate for user A should drop by If user A runs 10 streams and user B starts another 10, then transfer rate for user A should drop by 50% Common sense approach: Am I right? Not quite. Tomasz Wlodek DØRACE Workshop
When A uses 1 logical stream then user B with 10 streams will reduce his transfer rate from 1.4 to 1.05 Mbits/s When A uses 10 logical streams then appearance of user B with another 10 streams reduces his transfer rate from 7.5 to 5.3 Mbits/s Effect of “hooligan user” B with 10 logical streams Regardless whether A uses 1 or 10 streams, the “hooligan user” B will reduce the A transfer rate by ca 30%. This is very strange – and defies conventional wisdom too! Tomasz Wlodek DØRACE Workshop
Probably more than 1 “hooligan users” are needed to really bring the system down • Maybe with large number of users the transfer rate can be saturated. • Maybe one should consider intelligent bbftp: many users doing transfer simultaneously means: limit number of streams per user; few users doing transfer: allow more streams per user. Tomasz Wlodek DØRACE Workshop
Another parameter which could affect the transfer rate: buffer size. • Problem: bbftp version used in SAM is old and does not allow users to change the buffer size. • The most recent bbftp version is 2.2.0-beta0 and it allows change of buffer size. • I have installed bbftp 2.2.0-beta0 client on d0test and server on heppc13.uta.edu and did some tests Tomasz Wlodek DØRACE Workshop
Changing the buffer size Tomasz Wlodek DØRACE Workshop
Changing buffer size does not seem to have any obvious effect • Based on limited number of measurements • More parameters in bbftp to tune: “input and output window sizes” • currently investigating them Tomasz Wlodek DØRACE Workshop
How does SAM bbftp compare to bbftp-2.2.0-beta0 (the “cutting edge bbftp”) It is not possible to compare directly yet. SAM bbftp runs between hep.uta.edu and d0test Bbftp-2.2.0-beta0 runs between heppc13.uta.edu and d0test (I cannot install bbftp server on any computer I like) Tomasz Wlodek DØRACE Workshop
Conclusions • The parallel streams can increase transfer rate by a factor of 6-7 • The question “how it will affect other users” needs to be more carefully understood. • No reason to compress data during transfer – it does not help. • At this moment: changing buffer size does not improve transfer rate, but this needs more checking. • There are other parameters in bbftp which need to be understood (“window sizes”) • It is to early to say whether SAM should upgrade bbftp to most recent version. • A factor of 10 asymmetry in transfer to and from FNAL observed Tomasz Wlodek DØRACE Workshop