| layout | title |
|---|---|
page |
Differences between PI and PIE |
| Attr | PIE | PI |
|---|---|---|
| Drop Probability | p = alpha*(curr_q - ref_q) + beta(curr_q - old_q) + p | p = alpha(curr_q - ref_q) + beta(old_q - ref_q) + p |
| T_update | paper - 30ms, RFC - 16ms | paper - experiments section says sampling freq = 160 Hz which means 6.25 ms |
| alpha, beta | table of values depending on p | They are knobs that don't change with P |
| DQ_Threshold | paper - 10KB, RFC - 16 KB | - |
| ref_q_delay | paper - 20ms, RFC - 15ms | Experiments section says 200 packets |
| curr_q_delay | curr_q_len / avg_dep_rate | Not given |
| avg_dq_rate | avg_dq_rate = (1-w)avg_dq_rate + w(current_dq_rate) | - |
| w | paper - 0.5, RFC - 0.125 | - |
| current_dq_rate | current_dq_rate = dq_count / dq_time | - |
| burst_allowance | paper - 150ms, rfc - 144ms or 160ms | - |
The PI AQM works on using queue length and not queue delay. Hence components of PI involving Little's Law, calculation of average dequeue rate have been removed.