Hi,
I am looking at deterministic ways (perhaps employing data science) to
predict TCP throughput that i can expect between two end points. I am using
the latency (RTT) and the packet loss as the parameters. Is there anything
else that i can use to predict the throughput?
A related question to this is;
If i see an RTT of 150ms and packet loss of 0.01% between points A and B
and the maximum throughput then between these as, say 250Mbps. Then can i
say that i will *always* get the same (or in a close ballpark) throughput
not matter what time of the day i run these tests.
My points A and B can be virtual machines spawned on two different data
centers, say Amazon Virgina and Amazon Tokyo? So we're talking about long
distances here.
What else besides the RTT and packet loss can affect my TCP throughput
between two end points. I am assuming that the effects of a virtual machine
overload would have direct bearing on the RTT and packet loss, and hence
should cancel out. What i mean by this is that even if a VM is busy, then
that might induce larger losses and increased RTT, and that would affect my
TCP throughput. But then i already know what TCP throughput i get when i
have a given RTT and loss, and hence should be able to predict it.
Is there something that i am missing here?
Thanks, Glen
You need to account for window size as well. You should also account for
the details of the specific implementation of the TCP stack you are dealing
with if you truly need a deterministic result.
Only if you control the network load across the entire path.
As a simplified example, assume you did your test at 2AM and there's no
other activity, there's a bottleneck 1Gbps link in the path, and you get 250Mbps.
(yes, that result indicates probable misconfig, but bear with me.. 
You test again at 11AM, and now there's 7 other streams trying to pump 250Mbps
across that link. All 8 should probably drop back to 125Mbps.
For extra credit, factor in bufferbloat pushing your RTT through the roof
under congestion, and similar misbehaviors....
Oh, and that 0.01% packet loss is going to play heck with tcp slow-start
and opening the window - to quote RFC3649:
This document proposes HighSpeed TCP, a modification to TCP's
congestion control mechanism for use with TCP connections with large
congestion windows. In a steady-state environment, with a packet
loss rate p, the current Standard TCP's average congestion window is
roughly 1.2/sqrt(p) segments. This places a serious constraint on
the congestion windows that can be achieved by TCP in realistic
environments. For example, for a Standard TCP connection with 1500-
byte packets and a 100 ms round-trip time, achieving a steady-state
throughput of 10 Gbps would require an average congestion window of
83,333 segments, and a packet drop rate of at most one congestion
event every 5,000,000,000 packets (or equivalently, at most one
congestion event every 1 2/3 hours). The average packet drop rate of
at most 2*10^(-10) needed for full link utilization in this
environment corresponds to a bit error rate of at most 2*10^(-14),
and this is an unrealistic requirement for current networks.