Re: [PATCH v8 0/8] mm: Hot page tracking and promotion infrastructure - llama-bench

[Bharata B Rao]

llama-bench is the micro-benchmark shipped with llama.cpp. It loads a GGUF
model through the production inference path and reports two throughput numbers
per run: pp512 (one batched 512-token prefill, compute-bound) and tg128 (128
sequential autoregressive decode steps, bandwidth-bound).

Workload: llama-bench (llama.cpp), Mixtral-8x22B-Instruct Q4_K_M (140.6B
params, 79.7 GiB), -t 64 -p 512 -n 128 -r 5 --mmap 0.  pp512 = prefill
(compute-bound); tg128 = decode (memory-bandwidth-bound).

Setup: 3-node box, N0/N1 DRAM (256 GiB each), N2 CXL (distances 10/12/50).
Bench pinned to N1 CPUs with MPOL_PREFERRED_MANY({1}); a 200 GiB hot hog on
N1 forces kswapd to naturally demote ~1/3 of the model to CXL; each mode is
then measured while the hog keeps N1 under pressure.  pghot target_nid=1,
freq_threshold=1, freq_window=3000ms, rate_limit=65536 MBps, kmigrated
100ms/512.  hwhints runs arm AMD IBS (l3miss-only=1, period=10000).

Legend (columns)
----------------
  r1 = base / notier        7.2.0-rc2-base   nb=0 src=-
  r2 = base / tier (NUMAB2) 7.2.0-rc2-base   nb=2 src=-
  r3 = pghot / hintfaults   7.2.0-rc2-pghot  nb=2 src=0x1
  r4 = pghot / hwhints(IBS) 7.2.0-rc2-pghot  nb=0 src=0x2
  r5 = pghot / both         7.2.0-rc2-pghot  nb=2 src=0x3
  (nb = kernel.numa_balancing; src = pghot_enabled_sources)

Table 1 - Throughput (llama-bench, tokens/s over 5 reps)
-------------------------------------------------------
metric                       r1       r2       r3       r4       r5
-------------------------------------------------------------------
pp512 t/s                 69.31    62.46    54.53    53.95    51.49
  stddev                   2.70     6.65    10.41     8.12     4.79
tg128 t/s                 4.262    4.625    5.993    5.161    6.851
  stddev                  0.004    0.313    1.255    0.535    1.807
tg128 vs r1               1.00x    1.09x    1.41x    1.21x    1.61x
pp512 vs r1               1.00x    0.90x    0.79x    0.78x    0.74x

Table 1b - tg128 per-iteration (t/s), shows convergence
-------------------------------------------------------
iter                         r1       r2       r3       r4       r5
-------------------------------------------------------------------
iter1                     4.268    4.183    4.393    4.462    4.415
iter2                     4.261    4.457    5.336    4.820    5.602
iter3                     4.260    4.678    5.995    5.218    7.322
iter4                     4.261    4.841    6.501    5.494    8.152
iter5                     4.259    4.969    7.740    5.813    8.763

Table 2 - Key vmstat counters, run-phase deltas (millions of pages/events)
--------------------------------------------------------------------------
metric                          r1      r2      r3      r4      r5
------------------------------------------------------------------
pgpromote_success             0.00    3.05    6.95    2.17    6.53
pgpromote_candidate           0.00   14.34   24.12   13.83   21.09
pgdemote_kswapd               0.69    5.02   10.84    3.89    8.05
pgmigrate_success             0.69    8.07   17.80    6.06   14.59
numa_pte_updates              0.00   28.77   24.56    0.00   12.04
numa_hint_faults              0.00   27.77   24.12    0.00   11.39
pghot_recorded_accesses       0.00    0.00   24.12   13.83   21.28
pghot_recorded_hintfaults     0.00    0.00   24.12    0.00   11.39
pghot_recorded_hwhints        0.00    0.00    0.00   34.03   32.42
hwhint_total_events           0.00    0.00    0.00   34.03   32.42
hwhint_dram_accesses          0.00    0.00    0.00   20.03   22.34
hwhint_extmem_accesses        0.00    0.00    0.00   13.83    9.89

Key observations
----------------
1. pghot-both (both hintfaults and hwhints enabled) improves decode over both
   base baselines.
2. Prefill (pp512) suffers from tiering promotion (0.79x hintfaults, 0.78x hwhints,
   0.74x both vs notier); pghot/hf and /hw pp512 are noisy (stddev ~8-10).