Errata for Bounding SDRAM Interference: Detailed Analysis vs. Latency-Rate Analysis

Unfortunately, there was an error in Equation (5) in the paper, which made the improved bound for CCSP arbitration non-conservative. The problem was that the equation does not track fractional interference, i.e. the remainder (r) from the interference calculations (pi), between iterations. This prevents masters from accumulating fractional across iterations and hence fails to capture that over time, these fractions may add up to additional interference. Note that this is different from the pessimism in the original CCSP bound, where fractional potential from different masters are accumulated. This implies that the bound adds one unit of interference if two masters have enough potential for half a memory access each, which is different from the behavior of the actual arbiter as well as the improved bound presented in this paper.

The error in the equation does not affect any results shown in the paper. The correct bound that tracks remainders across iterations is shown below.