PCC context signatures: what an equal sum establishes ==================================================== The useful distinction --------------------- Sharing an I/O wrapper does not imply sharing its callers. Including caller addresses can distinguish contexts that a single wrapper address merges. Compressing those addresses into a sum still does not preserve full identity. The OSDI 2004 PCC design describes a 32-bit signature obtained by summing call-site PCs through the stack toward main(). The examples here use exact integer sums, without overflow, to isolate information lost by addition. Synthetic controls (hexadecimal) ------------------------------- Common list A: [0x1000, 0x2000, 0x4000], sum 0x7000. Identical list: same ordered frames, same sum. Different caller: [0x1000, 0x2000, 0x5000], sum 0x8000. Equal-sum list: [0x1000, 0x2800, 0x3800], sum 0x7000. Reordered callers: [0x1000, 0x4000, 0x2000], sum 0x7000. All cases share the first address. The different-caller control shows why using more context can help. The equal-sum case shows why a compressed key alone is not proof of context identity. The reordering control isolates the loss of sequence order. No overflow behavior is needed for either collision. Reproduce --------- Use Node.js with its built-in test runner. No third-party packages are needed. Keep model.mjs, model.test.mjs and examples.json together, then run: node model.mjs --check node --test model.test.mjs The tests include invalid inputs, unsigned-address boundaries and 4096 deterministic address lists checked against bit-column addition. This finite check validates the helper, not a collision probability or a storage system. Limits ------ Addresses are invented. These lists are not asserted to occur in any compiled program. The helper does not unwind native stacks, implement PCC's classifier, estimate reuse periods, model cache replacement or reproduce kernel results. Neither production incidence nor performance impact follows from these cases. A measured collision rate needs a declared workload and observed native contexts, including the exact signature implementation and address environment. Read alongside -------------- Gniady, Butt and Hu, Program-Counter-Based Pattern Classification in Buffer Caching, OSDI 2004, sections 4.1-4.3 and 5: https://www.usenix.org/legacy/publications/library/proceedings/osdi04/tech/full_papers/gnaidy/gnaidy_html/index.html Zhou, von Behren and Brewer, AMP: Program Context Specific Buffer Caching, USENIX ATC 2005. Its PCC* comparison retains a different cache manager: https://www.usenix.org/legacy/event/usenix05/tech/general/full_papers/short_papers/zhou/zhou_html/index.html