Network tomography enables network operators to detect and mitigate congestion by inferring internal states from multiple end-to-end active performance measurements. In previous studies, we proposed a novel architecture to develop P2P-based measurement system that scales with the nationwide Internet. One of the key features is low-overhead data management method where measurement results are piggybacked onto measurements themselves. However this feature leads to the loss of measurement results, which may cause inference accuracy to decline. In this paper, we estimate the proposed architecture from the viewpoint of tolerance to packet losses. Our simulation results show that the architecture can infer congested segments with satisfactory accuracy even if 1% of links in a network are congested.