mckee@MITRE.ARPA (H. Craig McKee) (10/29/86)
Bob Knight
Back on 25 Feb 86 you noted a large FTP load when new host tables were
released. The responses seemed to concentrate on what might be done
with the host tables. A more fundamental issue is the behavior of FTP
servers. I suggest that network performance would be improved if FTP
were revised as follows:
1. FTP should reject a request for a non-local transfer if "M"
incomming or "N" outgoing jobs are already in progress.
Rationale: When you permit many concurrent FTP jobs long connection
times are necessary and everyone gets slow service. Long connection
times increase the probability that equipment failure or a transitory
glitch will cause the connection to be aborted. In the absence of a
restart marker the user will have to start from the beginning of the
file. I estimate that sites with a single 56 Kbps PSN circuit should
limit concurrent FTP jobs to two incomming and two outgoing.
Once rejected, what's a user to do, sit around and harass the
FTP server? I suggest:
2. The FTP rejection message should allow the user to queue the
request for later execution.
Rationale: FTP jobs are not urgent. So long as the user knows the
job will eventually be completed, it doesn't matter if it takes several
hours.
Clark, Mills, and Nagle, among others, have offered good advice on
techniques to improve the performance of TCP and avoid network
congestion. The unstated, perhaps obvious, objective follows from
Queuing Theory: Never drive any path through the network at more than
70% of capacity. Being able to reject a request will smooth the FTP
load and should improve network throughput. Being able to queue jobs
will maintain the operational utility of FTP.
H. Craig McKee