Plan : staticdeps: enrich

plan/60_staticdeps.md

@@ -115,22 +115,36 @@ On SKX,
 
 #### With valgrind
 
+* Write a tool, valgrind-depsim, to instrument a binary to extract its
+  dependencies at runtime
+* Can extract memory, register and temp-based dependencies
+* Here, only the memory dependencies are relevant -- disable the other deps.
 * Instrument binary:
     * for each write, add `write_addr -> writer_pc` to a hashmap
     * for each read, fetch `writer_pc` from hashmap
         * if found, add a dependency `reader_pc -> writer_pc`
+    * use the process' memory map to translate PC to addresses inside ELF files
     * At the end, write deps file:
         * `#occur, src_elf_pc, src_elf_path, dst_elf_pc, dst_elf_path`
     * Run for each binary in genbenchs
+    * Takes about 1h on 30 parallel cores on Pinocchio; heavy memory usage
+
+Then, compare with staticdeps: `eval/vg_depsim.py` script.
 
 * For each binary in genbenchs,
+    * use genbench's bb split/occurrences to retrieve basic blocks
     * for each BB with more than 10% of max BB hits,
     * predict deps with staticdeps
-    * check which dependencies are found/missed from the instrumented ones
-    * limitation: will only find deps from/to the same BB! Dependencies leaving
-      a BB are discarded.
+        * cache the result: fast, but we're dealing with 3500 files.
+    * translate staticdeps' periodic deps to PC deps, discard the `iter`
+      parameter
+    * for each dependency from the depsim results that occurs inside this BB,
+        * check if found or missed, append to a list
+* score: `|found| / (|found| + |missed|)`. Discards occurrences.
+* limitation: will only find deps from/to the same BB! Dependencies leaving
+  a BB are discarded.
 
-* Result: about 38% of deps found.
+* Result: about 38% of deps found; 44% if weighting by occurrences
 
 * Cause: kernels executed in loops.
     * No dependency in the kernel
@@ -146,10 +160,41 @@ On SKX,
 * Fix: introduce dependency lifetime
     * timestamp = instructions executed (VG instrumentation, added up at the
       end of each BB)
-    * lifetime fixed to 1024 instructions
+    * lifetime fixed to 1024 instructions, order of magnitude of a ROB
     * dependencies are discarded if written to more than a lifetime ago
 
-* Result: about (?? TODO) of deps found
+* Result: about 58% of deps found; same if weighing.
+* If lifetime lowered to 512, about 56% of deps found, or 63% if weighing.
+    * Results are quite similar, lowering the lifetime further makes no
+      particular sense.
+
+Raw results:
+```
+In [123]: res_success(res_life512)
+Out[123]: 0.5640902544407105
+
+In [124]: res_success(res_life1024)
+Out[124]: 0.5761437608875034
+
+In [125]: res_success(res_nolife)
+Out[125]: 0.38143868803578085
+
+In [126]: res_success_weight(res_life512)
+Out[126]: 0.6347271857382266
+
+In [127]: res_success_weight(res_life1024)
+Out[127]: 0.5817404277466787
+
+In [128]: res_success_weight(res_nolife)
+Out[128]: 0.4397921976192802
+```
+
+* The results are reasonable, but not all the deps are caught
+* As argued above, will never see aliasing; important in plenty of cases.
+    * eg. if the compiler allocates `%rcx = A[i]` and `%rdx = A[i+2]` for some
+      reason, dependencies will be missed.
+* As argued in previous chapter, a complete dependencies analysis would require
+  a broader range: take the full scope into account
 
 #### With Gus
 
@@ -167,3 +212,9 @@ TODO ?
     * results
 * Run CesASMe on the no-memdeps suite with uiCA and uiCA+staticdeps
     * results
+
+* Although not all dependencies are detected [paragraph above], the "important"
+  ones seem to be detected: this is the most critical property for throughput
+  analysis
+    * but might not be true for other applications that require dependencies
+      detection