With the following experiments we will try to determine the effect of the number of columns on the standard error when estimating the proportion of dropped packets. We have considered three different scenarios, the first one has a smaller drop probability (10%), the second one has a large drop probability (60%) and the last one has a single row. As the figures below show, in all scenarios the percentile decreases with the number of columns, and it does so approximately proportional to the inverse of the square of the number of columns.

Small drop probability

Parameter Value
Packets 100
Drop probability 10%
Columns {8,16,32,64,128,256,512,1024}
Rows 32
Digest size 32
Hash function default
Xi function default
Pcap CAIDA-no dups
Average function mean

Large drop probability

Parameter Value
Packets 100
Drop probability 60%
Columns {8,16,32,64,128,256,512,1024}
Rows 32
Digest size 32
Hash function default
Xi function default
Pcap CAIDA-no dups
Average function mean

Single row sketch

Parameter Value
Packets 100
Drop probability 10%
Columns {8,16,32,64,128,256,512,1024}
Rows 1
Digest size 32
Hash function default
Xi function default
Pcap CAIDA-no dups
Average function mean