File size: 14,073 Bytes
aa74ad5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
# tiered-mem-traces

**60 memory access traces** for tiered-memory simulation, covering all 16
benchmark suites of the [CMU-SAFARI/DAMOV](https://github.com/CMU-SAFARI/DAMOV)
suite, plus **1 LLM-inference trace** collected from llama2.c (under
`llama_inference/`, Git LFS — see § LLM-inference trace).
Each trace preserves load/store/modify classification from
[Valgrind's lackey tool](https://valgrind.org/docs/manual/lk-manual.html).

## Format

Each trace is a gzipped text file with one memory operation per line:

```
S 1fff000198,8
S 1fff000190,8
S 1fff000188,8
L 04038e06,1
M 04038e10,4
```

- `L` = load, `S` = store, `M` = modify (read-modify-write)
- `<hex_addr>,<size>` = byte address and access size
- Instruction fetches are excluded; only data accesses are recorded

Total: 97 MB compressed (~4 GB uncompressed).

## Traces (60 total, by suite)

| Trace | Description | Access Pattern | Size |
|-------|-------------|----------------|-----:|
| **bwa** (genomics) | | | |
| `bwa.AlignLite` | BWA-MEM short-read alignment via FM-index lookup | scattered reads into a hot index region | 817K |
| **chai** (heterogeneous) | | | |
| `chai.BFS` | frontier-driven graph BFS over a CSR graph | scattered adjacency reads, drifting hot frontier | 1.1M |
| `chai.BS.BEZIER_KERNEL` | 2D Bezier surface evaluation | structured streaming with small reused control set | 1.1M |
| `chai.CEDD.Gaussian` | 5x5 Gaussian-blur image convolution | stencil streaming, near-perfect spatial locality | 1.3M |
| `chai.CEDD.Sobel` | 3x3 Sobel edge-detection | stencil streaming | 1.3M |
| `chai.HSTI` | input-side histogram increment | stream-read + scattered hot-write into bins | 782K |
| `chai.OOPPAD.OOPPAD` | out-of-place 2D padding (border insertion) | streaming with reflected-border copy | 1.3M |
| `chai.SC.SC` | stream compaction = filter + prefix-sum + scatter | two streaming passes + scattered scatter | 753K |
| **darknet** (ML) | | | |
| `darknet.conv` | im2col-style 2D convolution layer | weight streaming + activation reuse | 1.1M |
| **gase** (genomics) | | | |
| `gase.Align` | banded Smith-Waterman DP cell update | streaming along the band, three-row reuse | 4.1M |
| **hardware_effects** (microbench) | | | |
| `hardware_effects.bandwidth` | page-stride memory walk | scattered, anti-prefetcher pattern | 2.5M |
| `hardware_effects.false_sharing` | single-threaded RMW alternating two cache-line slots | high temporal locality, RMW writes | 259K |
| **hashjoin** (database) | | | |
| `hashjoin.MPSM` | massively-parallel sort-merge join | two-stream merge with run-length copies | 2.8M |
| `hashjoin.NPO.HashJoin` | no-partition open-addressed hash join (build + probe) | sequential probe + hot hash-table writes | 1.1M |
| `hashjoin.NPO_st` | single-thread no-partition variant of NPO | sequential probe + scattered hash-table reads | 2.8M |
| `hashjoin.PRO` | partitioned radix-style hash join | phase change between partition-build and probe | 1.5M |
| **hpcc** (HPC) | | | |
| `hpcc.RandomAccess` | GUPS-style xor-shift updates to random table slots | uniform-random scattered RMW | 2.5M |
| **hpcg** (HPC) | | | |
| `hpcg.CG` | conjugate-gradient inner loop on a 7-pt sparse stencil | SpMV scattered reads + vector streaming reductions | 2.2M |
| `hpcg.MG` | 3-level multigrid V-cycle (smooth + restrict + prolong) | stencil sweep at multiple resolutions | 408K |
| `hpcg.SpMV` | sparse matrix-vector multiply on a 27-pt stencil | row-streaming + scattered neighbor reads | 1.2M |
| **ligra** (graph) | | | |
| `ligra.BC` | betweenness centrality via Brandes-style two-pass BFS | frontier sweep + dependency back-propagation | 2.7M |
| `ligra.BellmanFord` | single-source shortest paths via repeated edge relaxation | edge-stream sweep with dist updates per round | 3.1M |
| `ligra.BFS` | frontier-driven BFS | per-level frontier sweep, drifting hot set | 1.9M |
| `ligra.Components` | connected-components via min-label propagation | full edge-stream sweep per round, label updates | 2.4M |
| `ligra.KCore` | k-core decomposition by iterative low-degree peeling | vertex-array sweep with neighbor degree updates | 2.0M |
| `ligra.MIS` | maximal independent set via Luby-style randomized priority | neighbor-priority compare + alive removal | 1.9M |
| `ligra.PageRank` | PageRank: full-vertex sweep with adjacency-list reads | per-iter full-vertex stream + scattered neighbor writes | 2.0M |
| `ligra.Triangle` | triangle counting via per-edge neighbor-set intersection | nested adjacency-list scans, heavy reads | 520K |
| **parboil** (HPC) | | | |
| `parboil.histo` | image histogram with ~16k bins | streaming read + scattered bin RMW | 884K |
| `parboil.mri-q` | MRI Q-matrix construction (per-voxel trig accumulation) | double-loop over voxels x k-points, streaming | 1.5M |
| `parboil.sgemm` | naive single-precision dense matmul | triple-loop with strided B access | 1.1M |
| `parboil.spmv` | sparse matrix-vector multiply (random sparsity) | row-stream + irregular column reads | 815K |
| **parsec** (multimedia) | | | |
| `parsec.Streamcluster` | k-means-style center-distance scan | stream over points + sequential per-DIM | 1.1M |
| **phoenix** (mapreduce) | | | |
| `phoenix.histogram` | 256-bin histogram over a streaming byte input | streaming read + dense hot-write into 256 bins | 863K |
| `phoenix.kmeans` | k-means iterations: assign + center update | phase change between assign and reduce | 963K |
| `phoenix.lr` | linear-regression gradient descent | streaming over (x, y) pairs | 956K |
| `phoenix.matrix-mult` | MapReduce-flavored dense matmul | row-by-row mapper-style sweep | 1.4M |
| `phoenix.reverseindex` | build a doc->terms reverse index | scattered hash-style writes + linked-list traversals | 744K |
| `phoenix.WordCount` | MapReduce word count via open-addressed hash table | streaming over words + scattered hash probes | 823K |
| **polybench** (HPC) | | | |
| `polybench.3mm` | three sequential dense matmuls (E=AB; F=CD; G=EF) | tiled streaming with phase change between matmuls | 1.6M |
| `polybench.atax` | y = A^T * (A x) (transpose + matrix-vector twice) | two streaming sweeps over A with vector reads | 897K |
| `polybench.bicg` | biconjugate-gradient kernel: q=Ap; s=A^T r | single sweep over A with two simultaneous reductions | 854K |
| `polybench.doitgen` | 3D tensor reduction | structured streaming with reduction | 1.2M |
| `polybench.gemm` | naive dense matmul C = A * B + C | triple-loop with strided B access | 1.4M |
| `polybench.syr2k` | symmetric rank-2 update C = AB^T + BA^T + C | triple-loop with shared inner-K accumulation | 1.6M |
| **rodinia** (heterogeneous) | | | |
| `rodinia.bfs` | BFS on a regular 4-connected 2D grid | regular per-level frontier sweep | 1.7M |
| `rodinia.cfd` | Euler-equation flux update on an unstructured mesh | edge-stream with scatter to two endpoints | 3.9M |
| `rodinia.hotspot` | 2D thermal-stencil sweep with two-array swap | 5-pt stencil, two-array swap per iter | 1.4M |
| `rodinia.lavamd` | short-range MD pairwise force calc with cutoff | O(N^2) pairwise with cutoff rejection | 1.4M |
| `rodinia.lud` | right-looking LU decomposition without pivoting | triangular sweep with shrinking trailing block | 1.6M |
| `rodinia.srad` | speckle-reducing anisotropic-diffusion image filter | two-pass per-pixel stencil | 1.8M |
| **splash2** (HPC) | | | |
| `splash2.FFT` | 1D FFT (bit-reversal + butterfly stages) | bit-reversal scatter + log(N) butterfly passes | 2.2M |
| `splash2.FMM` | O(N^2) particle pairwise interactions | O(N^2) double-loop with sqrt per pair | 966K |
| `splash2.OCEAN` | 2D 5-pt stencil sweep with two interleaved grids | stencil streaming with three-grid update | 1.9M |
| `splash2.RADIOSITY` | hierarchical 2D iterative form-factor refinement | KxK gather kernel applied per cell | 1.3M |
| `splash2.RAYTRACE` | ray casting through a small scene of spheres | per-pixel scene scan, frequent shared-data reads | 895K |
| **stream** (microbench) | | | |
| `stream.Add` | c[i] = a[i] + b[i] | perfectly sequential streaming, no reuse | 3.5M |
| `stream.Copy` | c[i] = a[i] | perfectly sequential streaming, no reuse | 3.5M |
| `stream.Scale` | b[i] = scalar * c[i] | perfectly sequential streaming, no reuse | 3.5M |
| `stream.Triad` | a[i] = b[i] + scalar * c[i] | perfectly sequential streaming, no reuse | 3.5M |

## LLM-inference trace

One extra trace lives under `llama_inference/` and is stored via **Git LFS**
(the file is ~323 MB compressed). It is **not** part of the 60-trace DAMOV
set above — different workload class, different scale.

| Trace | Description | Access pattern | Size |
|-------|-------------|----------------|-----:|
| `llama_inference/llama2c_stories15M_n3` | [karpathy/llama2.c](https://github.com/karpathy/llama2.c) inference, TinyStories-15M architecture (dim=288, hidden=768, 6 layers, vocab=32000), `n=3` forward passes (BOS prefill + 2 generated tokens, seed=42), random `N(0, 0.02)` weights — memory pattern is determined by tensor shapes and access order, not weight values, so it is structurally identical to a real TinyStories-15M checkpoint | continuous hot scratchpad (RunState) + linear weight-tensor sweeps + KV-cache growth per generated token | 323M |

Collection scale: 144,868,169 data accesses, 15,709 unique 4 KiB pages
(61.4 MiB footprint), captured under
`valgrind --tool=lackey --trace-mem=yes` with instruction fetches filtered
out. Loads 85.8 %, stores 13.7 %, modifies 0.4 %.

### Access pattern (annotated)

![llama2c stories15M, n=3 annotated VA](docs/llama_inference/llama2c_stories15M_n3_annotated.png)

The y-axis is the actual virtual address (stack at 131 GiB excluded — 3
pages absorbing 13.9 % of all accesses). Region boundaries are ground
truth from `/proc/self/maps` plus an instrumented `run.c` that printed
every weight / RunState / KV pointer under valgrind. Three forward passes
are visible as three structurally identical segments (dotted vertical
guides at ~46 M / ~84 M / ~122 M accesses):

- **64–66 MiB — HEAP / RunState scratchpad.** A continuous horizontal band
  touched throughout every pass. Holds `x`, `xb`, `xb2`, `q`, `att`,
  `logits` and other activation buffers. ~11 pages absorb ~50 % of the
  full trace's traffic; even with stack excluded, ~63 pages cover 50 % of
  the remaining accesses.
- **75–110 MiB — `wcls` / `token_embedding` classifier sweep.** The large
  diagonal climbing once per pass is the final `logits = matmul(x, wcls)`
  projection against the 32,000-entry vocabulary (`wcls` is shared with
  `token_embedding_table`). A 36.6 MiB linear scan at the end of every
  forward pass.
- **110–134 MiB — per-layer weight matmuls.** The "rainfall" pattern is
  six layers' worth of `wq`, `wk`, `wv`, `wo`, `w1`, `w2`, `w3` sweeps.
  Each cold weight page is touched once per forward pass.
- **133–137 MiB — KV cache.** `key_cache` + `value_cache` (3.4 MiB total).
  Pass 1 is pure stores; passes 2 and 3 have growing reads as the
  attention softmax iterates over an increasing number of KV slots.
- **131 GiB — stack (filtered in the figure).** 3 pages, 13.9 % of all
  accesses; valgrind redirects the user stack to the top of its managed
  address space.

The combination of a tiny continuously-hot scratchpad and a large
once-per-pass cold-weight sweep is the canonical heavy-tailed access
pattern of LLM inference: cache-policy choice barely matters at small
cache sizes because the hot set fits in any cache and the cold weight
scan is unaffected by eviction order.

### Reproducing

The collection pipeline (random-weight `model.bin` generator, valgrind
lackey wrapper, analyzer, plotter) lives in the arcsim repo:
[`scripts/llama_traces/`](https://github.com/alcriceedu/arcsim/tree/main/scripts/llama_traces).
See `scripts/llama_traces/README.md` for the full repro recipe and
`scripts/llama_traces/REPORT_stories15M.md` for the original
characterization run.

## How Traces Were Collected

Each trace comes from a small self-contained C facsimile that reproduces the
kernel's memory access pattern at smaller-than-production scale. The pipeline:

```
DAMOV recipe -> C facsimile (gcc -O0)
             -> valgrind --tool=lackey --trace-mem=yes
             -> awk filter (L/S/M lines only)
             -> gzip -6
             -> *.bytes.trace.gz
```

These are **pre-cache** traces (Valgrind sees every load/store including stack
traffic). Working sets are small (all 60 traces have U <= 3091 4-KiB pages).
The qualitative access pattern shape (skew, reuse-distance distribution,
hot-set drift, spatial locality class) is preserved from the original DAMOV
kernels, but absolute access counts differ from production-scale runs.

13 long-running workloads were SIGKILL-capped at ~25 min wall time once the
trace had grown to 25-200M accesses.

## Suite Coverage

| Suite | DAMOV Functions | Traces Here |
|-------|----------------:|------------:|
| bwa | 2 | 1 |
| chai | 19 | 7 |
| darknet | 4 | 1 |
| gase | 2 | 1 |
| hardware_effects | 4 | 2 |
| hashjoin | 12 | 4 |
| hpcc | 2 | 1 |
| hpcg | 5 | 3 |
| ligra | 39 | 8 |
| parboil | 7 | 4 |
| parsec | 5 | 1 |
| phoenix | 5 | 6 |
| polybench | 8 | 6 |
| rodinia | 11 | 6 |
| splash2 | 14 | 5 |
| stream | 4 | 4 |
| **Total** | **143** | **60** |

## Attribution

These traces are derived from the **DAMOV** benchmark suite by Oliveira et al.
(CMU-SAFARI). The kernels were re-implemented at small scale as self-contained
C facsimiles; the intent and naming of each trace follows the original DAMOV
classification.

> Geraldo F. Oliveira, Juan Gomez-Luna, Lois Orosa, Saugata Ghose,
> Nandita Vijaykumar, Ivan Fernandez, Mohammad Sadrosadati, Onur Mutlu.
> "DAMOV: A New Methodology and Benchmark Suite for Evaluating Data
> Movement Bottlenecks." IEEE Access, 2021. arXiv:2105.03725
>
> https://github.com/CMU-SAFARI/DAMOV

DAMOV aggregates kernels from BWA, Chai, Darknet, GASE, Hardware-Effects,
HPC-Challenge, HPCG, Hashjoin, Ligra, PARSEC, Parboil, PolyBench, Phoenix,
Rodinia, SPLASH-2, and STREAM -- each licensed and credited by its upstream
authors.