Prefetch: capacity-mode readahead, plus the model and accuracy experiment

CMakeLists.txt

@@ -57,6 +57,7 @@ add_library(
     src/cuda_runtime/timing_binding.cpp
     src/cuda_runtime/vmm.cpp
     src/hybrid/calibrator.cpp
+    src/prefetch/prefetch_model.cpp
     src/profile/profile.cpp
     src/host_service/request_dispatcher.cpp
     src/host_service/backing_store.cpp
@@ -108,6 +109,19 @@ target_link_libraries(hbfsim_profile_tests PRIVATE hbfsim_core)
 add_test(NAME profile COMMAND hbfsim_profile_tests)
 set_tests_properties(profile PROPERTIES WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}")
 
+add_executable(hbf_prefetch_bench benchmarks/prefetch/hbf_prefetch_bench.cpp)
+target_link_libraries(hbf_prefetch_bench PRIVATE hbfsim_core)
+
+add_executable(
+    hbf_capacity_readahead_bench
+    benchmarks/prefetch/hbf_capacity_readahead_bench.cpp
+)
+target_link_libraries(hbf_capacity_readahead_bench PRIVATE hbfsim_core)
+
+add_executable(prefetch_model_test tests/cpu/prefetch_model_test.cpp)
+target_link_libraries(prefetch_model_test PRIVATE hbfsim_core)
+add_test(NAME prefetch_model COMMAND prefetch_model_test)
+
 add_executable(hbfsim_calibrator_tests tests/cpu/calibrator_test.cpp)
 target_link_libraries(hbfsim_calibrator_tests PRIVATE hbfsim_core)
 add_test(NAME calibrator COMMAND hbfsim_calibrator_tests)
@@ -167,6 +181,12 @@ add_test(
     NAME capacity_page_service COMMAND hbfsim_capacity_page_service_tests
 )
 
+add_executable(
+    hbfsim_capacity_readahead_tests tests/cpu/capacity_readahead_test.cpp
+)
+target_link_libraries(hbfsim_capacity_readahead_tests PRIVATE hbfsim_core)
+add_test(NAME capacity_readahead COMMAND hbfsim_capacity_readahead_tests)
+
 add_executable(
     hbfsim_capacity_handoff_tests tests/cpu/capacity_handoff_test.cpp
 )

benchmarks/prefetch/README.md

@@ -0,0 +1,111 @@
+# Prefetch experiment: what to run and where everything is
+
+The reasoning behind the experiment, and what its numbers may and may not be
+used to claim, is in `docs/46-预取实验设计.md`. This file is only the map and
+the commands.
+
+## Files
+
+| Path | What it is |
+|---|---|
+| `include/hbfsim/prefetch_model.hpp` | The model's interface, with each knob's meaning |
+| `src/prefetch/prefetch_model.cpp` | The model: a deterministic discrete-event simulation |
+| `tests/cpu/prefetch_model_test.cpp` | 12 property assertions, written before the implementation |
+| `benchmarks/prefetch/hbf_prefetch_bench.cpp` | Sweeps one access stream through the model and prints JSON |
+| `src/host_service/capacity_page_service.cpp` | The real system-side readahead, in capacity mode |
+| `tests/cpu/capacity_readahead_test.cpp` | Its contract: off by default, never forces a writeback, never drains on the demand's own path |
+| `benchmarks/prefetch/hbf_capacity_readahead_bench.cpp` | Drives the real readahead and reports media reads avoided |
+| `scripts/run_prefetch_accuracy_sweep.py` | Runs all three streams and writes the artifact and a CSV |
+| `docs/proofs/artifacts/prefetch-accuracy-sweep.json` | The swept cells |
+| `docs/proofs/artifacts/prefetch-accuracy-sweep.csv` | The same cells, flat, for plotting |
+
+## Build
+
+The model and its test are CPU-only. Neither needs a GPU or `nvcc`.
+
+```
+cmake -S . -B build -DHBFSIM_ENABLE_CUDA=OFF -DHBFSIM_ENABLE_MQSIM=ON \
+      -DCMAKE_BUILD_TYPE=Release
+cmake --build build -j"$(nproc)" --target prefetch_model_test hbf_prefetch_bench
+```
+
+## Run the tests
+
+```
+./build/prefetch_model_test    # exit code 0, prints nothing on success
+```
+
+On failure it prints the line number and the assertion that failed.
+
+## Reproduce the sweep
+
+```
+python3 scripts/run_prefetch_accuracy_sweep.py --build-dir build
+```
+
+Writes `docs/proofs/artifacts/prefetch-accuracy-sweep.json` and the matching
+`.csv`. Both carry a `disclaimer` field reading
+`modeled, not measured on any device or GPU`.
+
+## One stream at a time
+
+```
+./build/hbf_prefetch_bench --stream sequential --accesses 20000
+./build/hbf_prefetch_bench --stream random     --accesses 20000
+./build/hbf_prefetch_bench --stream moe        --accesses 20000 \
+                           --pages-per-expert 2304
+```
+
+`--pages-per-expert 2304` is one Qwen3-30B-A3B expert: 3 x 2048 x 768
+parameters in bf16 is 9,437,184 bytes, which is 2304 pages of 4 KiB.
+
+Other options: `--compute-ns` (accelerator time between two accesses, the
+interval a prefetch hides behind), `--lead` (how many accesses ahead a prefetch
+is issued), `--buffer-pages`, `--max-in-flight`, `--seed`.
+
+## The real readahead, as opposed to the model
+
+`hbf_capacity_readahead_bench` exercises `CapacityPageService` directly. It
+reports media reads avoided, not wall-clock time, and touches no GPU.
+
+Does the implementation reproduce what the model predicts for a next-page
+policy, which is (P-1)/P for P pages per expert:
+
+```
+for p in 4 8 16; do
+  ./build/hbf_capacity_readahead_bench --stream moe --pages-per-expert $p \
+      --readahead $p --frames 256 --drain-per-demand 0
+done
+```
+
+What the worker's drain rate is worth. `--drain-per-demand` is how many queued
+pages the worker gets through between two demands, and 0 means it keeps up
+completely:
+
+```
+for d in 0 4 2 1; do
+  ./build/hbf_capacity_readahead_bench --stream moe --pages-per-expert 8 \
+      --readahead 8 --frames 256 --drain-per-demand $d
+done
+```
+
+Draining one page per demand does not merely reduce the benefit. It turns an
+86.79% reduction in media reads into a 14.17% increase, because queued pages go
+stale before they are used while still costing bandwidth and frames.
+
+## Reproducing the two results the paper leans on
+
+The naive next-page policy's accuracy on a Mixture-of-Experts stream is
+(P-1)/P, where P is how many pages one expert occupies:
+
+```
+for p in 1 2 4 8 16; do
+  ./build/hbf_prefetch_bench --stream moe --accesses 20000 --pages-per-expert $p
+done
+```
+
+One token of Qwen3-30B-A3B, which is 884,736 page accesses:
+
+```
+./build/hbf_prefetch_bench --stream moe --accesses 20000 --pages-per-expert 2304
+```