2025-11-07 01:27:04 +09:00
|
|
|
// pool_api.inc.h — Box: L2 Pool public API (alloc/free/lookup)
|
|
|
|
|
#ifndef POOL_API_INC_H
|
|
|
|
|
#define POOL_API_INC_H
|
|
|
|
|
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
#include "pagefault_telemetry_box.h" // Box PageFaultTelemetry (PF_BUCKET_MID)
|
|
|
|
|
|
2025-11-07 01:27:04 +09:00
|
|
|
void* hak_pool_try_alloc(size_t size, uintptr_t site_id) {
|
2025-11-09 18:55:50 +09:00
|
|
|
// Debug: IMMEDIATE output to verify function is called
|
|
|
|
|
static int first_call = 1;
|
2025-11-11 01:47:06 +09:00
|
|
|
if (first_call) { HAKMEM_LOG("[Pool] hak_pool_try_alloc FIRST CALL EVER!\n"); first_call = 0; }
|
2025-11-09 18:55:50 +09:00
|
|
|
|
2025-11-11 01:47:06 +09:00
|
|
|
if (size == 40960) { HAKMEM_LOG("[Pool] hak_pool_try_alloc called with 40KB (Bridge class 5)\n"); }
|
2025-11-09 18:55:50 +09:00
|
|
|
|
2025-11-07 01:27:04 +09:00
|
|
|
hak_pool_init(); // pthread_once() ensures thread-safe init (no data race!)
|
2025-11-09 18:55:50 +09:00
|
|
|
|
|
|
|
|
// Debug for 33-41KB allocations
|
2025-11-11 01:47:06 +09:00
|
|
|
if (size >= 33000 && size <= 41000) { HAKMEM_LOG("[Pool] hak_pool_try_alloc: size=%zu (after init)\n", size); }
|
2025-11-09 18:55:50 +09:00
|
|
|
|
2025-11-07 01:27:04 +09:00
|
|
|
// P1.7 approach: Avoid using pool during ALL wrapper calls (conservative but safe)
|
|
|
|
|
extern int hak_in_wrapper(void);
|
2025-11-09 18:55:50 +09:00
|
|
|
if (hak_in_wrapper() && !g_wrap_l2_enabled) {
|
2025-11-11 01:47:06 +09:00
|
|
|
if (size >= 33000 && size <= 41000) { HAKMEM_LOG("[Pool] REJECTED: in_wrapper=%d, wrap_l2=%d\n", hak_in_wrapper(), g_wrap_l2_enabled); }
|
2025-11-09 18:55:50 +09:00
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
if (!hak_pool_is_poolable(size)) {
|
2025-11-11 01:47:06 +09:00
|
|
|
if (size >= 33000 && size <= 41000) { HAKMEM_LOG("[Pool] REJECTED: not poolable (min=%d, max=%d)\n", POOL_MIN_SIZE, POOL_MAX_SIZE); }
|
2025-11-09 18:55:50 +09:00
|
|
|
return NULL;
|
|
|
|
|
}
|
2025-11-07 01:27:04 +09:00
|
|
|
|
|
|
|
|
// Get class and shard indices
|
|
|
|
|
int class_idx = hak_pool_get_class_index(size);
|
2025-11-09 18:55:50 +09:00
|
|
|
if (class_idx < 0) {
|
2025-11-11 01:47:06 +09:00
|
|
|
if (size >= 33000 && size <= 41000) { HAKMEM_LOG("[Pool] REJECTED: class_idx=%d (size=%zu not mapped)\n", class_idx, size); }
|
2025-11-09 18:55:50 +09:00
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
|
2025-11-11 01:47:06 +09:00
|
|
|
if (size >= 33000 && size <= 41000) { HAKMEM_LOG("[Pool] ACCEPTED: class_idx=%d, proceeding with allocation\n", class_idx); }
|
2025-11-07 01:27:04 +09:00
|
|
|
|
|
|
|
|
// MF2: Per-Page Sharding path
|
|
|
|
|
if (g_mf2_enabled) {
|
|
|
|
|
return mf2_alloc_fast(class_idx, size, site_id);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// OLD PATH: TLS fast path (ring then local LIFO); drain TC only when needed
|
|
|
|
|
PoolTLSRing* ring = &g_tls_bin[class_idx].ring;
|
|
|
|
|
if (g_tc_enabled && ring->top < g_tc_drain_trigger && mid_tc_has_items(class_idx)) {
|
|
|
|
|
HKM_TIME_START(t_tc_drain);
|
|
|
|
|
if (mid_tc_drain_into_tls(class_idx, ring, &g_tls_bin[class_idx])) {
|
|
|
|
|
HKM_TIME_END(HKM_CAT_TC_DRAIN, t_tc_drain);
|
|
|
|
|
if (ring->top > 0) {
|
|
|
|
|
HKM_TIME_START(t_ring_pop0);
|
|
|
|
|
PoolBlock* tlsb = ring->items[--ring->top];
|
|
|
|
|
HKM_TIME_END(HKM_CAT_POOL_TLS_RING_POP, t_ring_pop0);
|
|
|
|
|
void* raw = (void*)tlsb;
|
|
|
|
|
AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
void* user0 = (char*)raw + HEADER_SIZE;
|
2025-11-07 01:27:04 +09:00
|
|
|
mid_page_inuse_inc(raw);
|
|
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
|
|
|
|
|
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
pagefault_telemetry_touch(PF_BUCKET_MID, user0);
|
|
|
|
|
return user0;
|
2025-11-07 01:27:04 +09:00
|
|
|
}
|
|
|
|
|
} else { HKM_TIME_END(HKM_CAT_TC_DRAIN, t_tc_drain); }
|
|
|
|
|
}
|
|
|
|
|
if (g_tls_ring_enabled) {
|
|
|
|
|
if (ring->top == 0) {
|
|
|
|
|
atomic_fetch_add_explicit(&g_pool.ring_underflow, 1, memory_order_relaxed);
|
|
|
|
|
}
|
|
|
|
|
if (ring->top > 0) {
|
|
|
|
|
HKM_TIME_START(t_ring_pop1);
|
|
|
|
|
PoolBlock* tlsb = ring->items[--ring->top];
|
|
|
|
|
HKM_TIME_END(HKM_CAT_POOL_TLS_RING_POP, t_ring_pop1);
|
|
|
|
|
void* raw = (void*)tlsb;
|
|
|
|
|
AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
void* user1 = (char*)raw + HEADER_SIZE;
|
2025-11-07 01:27:04 +09:00
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
|
|
|
|
|
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
pagefault_telemetry_touch(PF_BUCKET_MID, user1);
|
|
|
|
|
return user1;
|
2025-11-07 01:27:04 +09:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (g_tls_bin[class_idx].lo_head) {
|
|
|
|
|
HKM_TIME_START(t_lifo_pop0);
|
|
|
|
|
PoolBlock* b = g_tls_bin[class_idx].lo_head;
|
|
|
|
|
g_tls_bin[class_idx].lo_head = b->next;
|
|
|
|
|
if (g_tls_bin[class_idx].lo_count) g_tls_bin[class_idx].lo_count--;
|
|
|
|
|
HKM_TIME_END(HKM_CAT_POOL_TLS_LIFO_POP, t_lifo_pop0);
|
|
|
|
|
void* raw = (void*)b; AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
void* user2 = (char*)raw + HEADER_SIZE;
|
2025-11-07 01:27:04 +09:00
|
|
|
mid_page_inuse_inc(raw);
|
|
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
|
|
|
|
|
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
pagefault_telemetry_touch(PF_BUCKET_MID, user2);
|
|
|
|
|
return user2;
|
2025-11-07 01:27:04 +09:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Compute shard only when we need to access shared structures
|
|
|
|
|
int shard_idx = hak_pool_get_shard_index(site_id);
|
|
|
|
|
|
|
|
|
|
// Try to batch-pop from a non-empty shard using trylock to fill TLS ring
|
|
|
|
|
if (g_tls_ring_enabled) {
|
|
|
|
|
int s0 = choose_nonempty_shard(class_idx, shard_idx);
|
|
|
|
|
for (int probe = 0; probe < g_trylock_probes; ++probe) {
|
|
|
|
|
int s = (s0 + probe) & (POOL_NUM_SHARDS - 1);
|
|
|
|
|
pthread_mutex_t* l = &g_pool.freelist_locks[class_idx][s].m;
|
|
|
|
|
atomic_fetch_add_explicit(&g_pool.trylock_attempts, 1, memory_order_relaxed);
|
|
|
|
|
if (pthread_mutex_trylock(l) == 0) {
|
|
|
|
|
atomic_fetch_add_explicit(&g_pool.trylock_success, 1, memory_order_relaxed);
|
|
|
|
|
// First, drain any remote frees into freelist
|
|
|
|
|
if (atomic_load_explicit(&g_pool.remote_count[class_idx][s], memory_order_relaxed) != 0) {
|
|
|
|
|
drain_remote_locked(class_idx, s);
|
|
|
|
|
}
|
|
|
|
|
PoolBlock* head = g_pool.freelist[class_idx][s];
|
|
|
|
|
int to_ring = POOL_L2_RING_CAP - ring->top; if (to_ring < 0) to_ring = 0;
|
|
|
|
|
while (head && to_ring-- > 0) { PoolBlock* nxt = head->next; ring->items[ring->top++] = head; head = nxt; }
|
|
|
|
|
while (head) { PoolBlock* nxt = head->next; head->next = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = head; g_tls_bin[class_idx].lo_count++; head = nxt; }
|
|
|
|
|
g_pool.freelist[class_idx][s] = head;
|
|
|
|
|
if (!head) clear_nonempty_bit(class_idx, s);
|
|
|
|
|
pthread_mutex_unlock(l);
|
|
|
|
|
if (ring->top > 0) {
|
|
|
|
|
PoolBlock* tlsb = ring->items[--ring->top];
|
|
|
|
|
void* raw = (void*)tlsb;
|
|
|
|
|
AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
|
|
|
|
|
mid_page_inuse_inc(raw);
|
|
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
|
|
|
|
|
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
|
|
|
|
|
return (char*)raw + HEADER_SIZE;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Try TLS active pages (owner-only local bump-run, up to 3)
|
|
|
|
|
PoolTLSPage* ap = NULL;
|
|
|
|
|
if (g_tls_active_page_a[class_idx].page && g_tls_active_page_a[class_idx].count > 0 && g_tls_active_page_a[class_idx].bump < g_tls_active_page_a[class_idx].end) ap = &g_tls_active_page_a[class_idx];
|
|
|
|
|
else if (g_tls_active_page_b[class_idx].page && g_tls_active_page_b[class_idx].count > 0 && g_tls_active_page_b[class_idx].bump < g_tls_active_page_b[class_idx].end) ap = &g_tls_active_page_b[class_idx];
|
|
|
|
|
else if (g_tls_active_page_c[class_idx].page && g_tls_active_page_c[class_idx].count > 0 && g_tls_active_page_c[class_idx].bump < g_tls_active_page_c[class_idx].end) ap = &g_tls_active_page_c[class_idx];
|
|
|
|
|
if (ap) {
|
|
|
|
|
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) {
|
|
|
|
|
int need = POOL_L2_RING_CAP - ring->top;
|
|
|
|
|
(void)refill_tls_from_active_page(class_idx, ring, &g_tls_bin[class_idx], ap, need);
|
|
|
|
|
}
|
|
|
|
|
PoolBlock* b = NULL;
|
|
|
|
|
if (ring->top > 0) { b = ring->items[--ring->top]; }
|
|
|
|
|
else if (ap->page && ap->count > 0 && ap->bump < ap->end) {
|
|
|
|
|
b = (PoolBlock*)(void*)ap->bump; ap->bump += (HEADER_SIZE + g_class_sizes[class_idx]); ap->count--; if (ap->bump >= ap->end || ap->count<=0){ ap->page=NULL; ap->count=0; }
|
|
|
|
|
}
|
|
|
|
|
if (b) {
|
|
|
|
|
void* raw = (void*)b; AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
|
|
|
|
|
mid_page_inuse_inc(raw);
|
|
|
|
|
g_pool.hits[class_idx]++;
|
|
|
|
|
return (char*)raw + HEADER_SIZE;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lock the shard freelist for this (class, shard)
|
|
|
|
|
pthread_mutex_t* lock = &g_pool.freelist_locks[class_idx][shard_idx].m;
|
|
|
|
|
HKM_TIME_START(t_lock);
|
|
|
|
|
struct timespec ts_lk1; int lk1 = hkm_prof_begin(&ts_lk1);
|
|
|
|
|
(void)ts_lk1; (void)lk1; // Unused profiling variables
|
|
|
|
|
pthread_mutex_lock(lock);
|
|
|
|
|
HKM_TIME_END(HKM_CAT_POOL_LOCK, t_lock);
|
|
|
|
|
hkm_prof_end(lk1, HKP_POOL_LOCK, &ts_lk1);
|
|
|
|
|
|
|
|
|
|
// Try to pop from freelist
|
|
|
|
|
PoolBlock* block = g_pool.freelist[class_idx][shard_idx];
|
|
|
|
|
|
|
|
|
|
if (!block) {
|
|
|
|
|
// Before refilling, try draining remote stack and simple shard steal
|
|
|
|
|
int stole = 0;
|
|
|
|
|
const FrozenPolicy* pol = hkm_policy_get();
|
|
|
|
|
if (pol) {
|
|
|
|
|
uint16_t cap = 0;
|
|
|
|
|
if (class_idx < 5) cap = pol->mid_cap[class_idx];
|
|
|
|
|
else if (class_idx == 5 && pol->mid_dyn1_bytes != 0) cap = pol->mid_cap_dyn1;
|
|
|
|
|
else if (class_idx == 6 && pol->mid_dyn2_bytes != 0) cap = pol->mid_cap_dyn2;
|
|
|
|
|
// Drain remotes
|
|
|
|
|
if (atomic_load_explicit(&g_pool.remote_count[class_idx][shard_idx], memory_order_relaxed) != 0) {
|
|
|
|
|
drain_remote_locked(class_idx, shard_idx);
|
|
|
|
|
block = g_pool.freelist[class_idx][shard_idx];
|
|
|
|
|
}
|
|
|
|
|
// Light shard steal when over cap
|
|
|
|
|
if (!block && cap > 0 && g_pool.pages_by_class[class_idx] >= cap) {
|
|
|
|
|
HKM_TIME_START(t_steal);
|
|
|
|
|
for (int d = 1; d <= 4 && !stole; d++) {
|
|
|
|
|
int s1 = (shard_idx + d) & (POOL_NUM_SHARDS - 1);
|
|
|
|
|
int s2 = (shard_idx - d) & (POOL_NUM_SHARDS - 1);
|
|
|
|
|
if (is_shard_nonempty(class_idx, s1)) {
|
|
|
|
|
pthread_mutex_t* l2 = &g_pool.freelist_locks[class_idx][s1].m;
|
|
|
|
|
pthread_mutex_lock(l2);
|
|
|
|
|
PoolBlock* b2 = g_pool.freelist[class_idx][s1];
|
|
|
|
|
if (b2) {
|
|
|
|
|
g_pool.freelist[class_idx][s1] = b2->next;
|
|
|
|
|
if (!g_pool.freelist[class_idx][s1]) clear_nonempty_bit(class_idx, s1);
|
|
|
|
|
block = b2; stole = 1;
|
|
|
|
|
}
|
|
|
|
|
pthread_mutex_unlock(l2);
|
|
|
|
|
}
|
|
|
|
|
if (!stole && is_shard_nonempty(class_idx, s2)) {
|
|
|
|
|
pthread_mutex_t* l3 = &g_pool.freelist_locks[class_idx][s2].m;
|
|
|
|
|
pthread_mutex_lock(l3);
|
|
|
|
|
PoolBlock* b3 = g_pool.freelist[class_idx][s2];
|
|
|
|
|
if (b3) {
|
|
|
|
|
g_pool.freelist[class_idx][s2] = b3->next;
|
|
|
|
|
if (!g_pool.freelist[class_idx][s2]) clear_nonempty_bit(class_idx, s2);
|
|
|
|
|
block = b3; stole = 1;
|
|
|
|
|
}
|
|
|
|
|
pthread_mutex_unlock(l3);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
HKM_TIME_END(HKM_CAT_SHARD_STEAL, t_steal);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (!stole && !block) {
|
|
|
|
|
// Freelist empty, refill page
|
|
|
|
|
PoolTLSPage* tap = NULL;
|
|
|
|
|
if (g_tls_active_page_a[class_idx].page == NULL || g_tls_active_page_a[class_idx].count == 0) tap = &g_tls_active_page_a[class_idx];
|
|
|
|
|
else if (g_tls_active_page_b[class_idx].page == NULL || g_tls_active_page_b[class_idx].count == 0) tap = &g_tls_active_page_b[class_idx];
|
|
|
|
|
else if (g_tls_active_page_c[class_idx].page == NULL || g_tls_active_page_c[class_idx].count == 0) tap = &g_tls_active_page_c[class_idx];
|
|
|
|
|
else tap = &g_tls_active_page_a[class_idx];
|
|
|
|
|
HKM_TIME_START(t_alloc_page);
|
|
|
|
|
if (alloc_tls_page(class_idx, tap)) {
|
|
|
|
|
HKM_TIME_END(HKM_CAT_POOL_ALLOC_TLS_PAGE, t_alloc_page);
|
|
|
|
|
pthread_mutex_unlock(lock);
|
|
|
|
|
// Top-up ring and return
|
|
|
|
|
ap = tap;
|
|
|
|
|
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) {
|
|
|
|
|
int need = POOL_L2_RING_CAP - ring->top;
|
|
|
|
|
(void)refill_tls_from_active_page(class_idx, ring, &g_tls_bin[class_idx], ap, need);
|
|
|
|
|
}
|
|
|
|
|
PoolBlock* takeb = NULL;
|
|
|
|
|
if (ring->top > 0) { HKM_TIME_START(t_ring_pop2); takeb = ring->items[--ring->top]; HKM_TIME_END(HKM_CAT_POOL_TLS_RING_POP, t_ring_pop2);}
|
|
|
|
|
else if (ap->page && ap->count > 0 && ap->bump < ap->end) { takeb = (PoolBlock*)(void*)ap->bump; ap->bump += (HEADER_SIZE + g_class_sizes[class_idx]); ap->count--; if (ap->bump >= ap->end || ap->count==0){ ap->page=NULL; ap->count=0; } }
|
|
|
|
|
void* raw2 = (void*)takeb; AllocHeader* hdr2 = (AllocHeader*)raw2;
|
|
|
|
|
mid_set_header(hdr2, g_class_sizes[class_idx], site_id);
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
void* user3 = (char*)raw2 + HEADER_SIZE;
|
2025-11-07 01:27:04 +09:00
|
|
|
mid_page_inuse_inc(raw2);
|
|
|
|
|
g_pool.hits[class_idx]++;
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
pagefault_telemetry_touch(PF_BUCKET_MID, user3);
|
|
|
|
|
return user3;
|
2025-11-07 01:27:04 +09:00
|
|
|
}
|
|
|
|
|
HKM_TIME_START(t_refill);
|
|
|
|
|
struct timespec ts_rf; int rf = hkm_prof_begin(&ts_rf);
|
|
|
|
|
int ok = refill_freelist(class_idx, shard_idx);
|
|
|
|
|
HKM_TIME_END(HKM_CAT_POOL_REFILL, t_refill);
|
|
|
|
|
hkm_prof_end(rf, HKP_POOL_REFILL, &ts_rf);
|
|
|
|
|
if (!ok) {
|
|
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
|
|
|
|
|
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.misses[class_idx]++;
|
|
|
|
|
pthread_mutex_unlock(lock);
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Pop block and adopt page
|
|
|
|
|
g_pool.freelist[class_idx][shard_idx] = block->next;
|
|
|
|
|
mid_desc_adopt(block, class_idx, (uint64_t)(uintptr_t)pthread_self());
|
|
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5;
|
|
|
|
|
if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.hits[class_idx]++;
|
|
|
|
|
if (g_pool.freelist[class_idx][shard_idx] == NULL) clear_nonempty_bit(class_idx, shard_idx);
|
|
|
|
|
pthread_mutex_unlock(lock);
|
|
|
|
|
|
|
|
|
|
// Store to TLS then pop
|
|
|
|
|
PoolBlock* take;
|
|
|
|
|
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) { ring->items[ring->top++] = block; take = ring->items[--ring->top]; }
|
|
|
|
|
else { block->next = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = block; g_tls_bin[class_idx].lo_count++;
|
|
|
|
|
if (g_tls_ring_enabled && ring->top > 0) { take = ring->items[--ring->top]; }
|
|
|
|
|
else { take = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = take->next; if (g_tls_bin[class_idx].lo_count) g_tls_bin[class_idx].lo_count--; } }
|
|
|
|
|
|
|
|
|
|
void* raw = (void*)take; AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
mid_set_header(hdr, g_class_sizes[class_idx], site_id);
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
void* user4 = (char*)raw + HEADER_SIZE;
|
2025-11-07 01:27:04 +09:00
|
|
|
mid_page_inuse_inc(raw);
|
Phase 23 Unified Cache + PageFaultTelemetry generalization: Mid/VM page-fault bottleneck identified
Summary:
- Phase 23 Unified Cache: +30% improvement (Random Mixed 256B: 18.18M → 23.68M ops/s)
- PageFaultTelemetry: Extended to generic buckets (C0-C7, MID, L25, SSM)
- Measurement-driven decision: Mid/VM page-faults (80-100K) >> Tiny (6K) → prioritize Mid/VM optimization
Phase 23 Changes:
1. Unified Cache implementation (core/front/tiny_unified_cache.{c,h})
- Direct SuperSlab carve (TLS SLL bypass)
- Self-contained pop-or-refill pattern
- ENV: HAKMEM_TINY_UNIFIED_CACHE=1, HAKMEM_TINY_UNIFIED_C{0-7}=128
2. Fast path pruning (tiny_alloc_fast.inc.h, tiny_free_fast_v2.inc.h)
- Unified ON → direct cache access (skip all intermediate layers)
- Alloc: unified_cache_pop_or_refill() → immediate fail to slow
- Free: unified_cache_push() → fallback to SLL only if full
PageFaultTelemetry Changes:
3. Generic bucket architecture (core/box/pagefault_telemetry_box.{c,h})
- PF_BUCKET_{C0-C7, MID, L25, SSM} for domain-specific measurement
- Integration: hak_pool_try_alloc(), l25_alloc_new_run(), shared_pool_allocate_superslab_unlocked()
4. Measurement results (Random Mixed 500K / 256B):
- Tiny C2-C7: 2-33 pages, high reuse (64-3.8 touches/page)
- SSM: 512 pages (initialization footprint)
- MID/L25: 0 (unused in this workload)
- Mid/Large VM benchmarks: 80-100K page-faults (13-16x higher than Tiny)
Ring Cache Enhancements:
5. Hot Ring Cache (core/front/tiny_ring_cache.{c,h})
- ENV: HAKMEM_TINY_HOT_RING_ENABLE=1, HAKMEM_TINY_HOT_RING_C{0-7}=size
- Conditional compilation cleanup
Documentation:
6. Analysis reports
- RANDOM_MIXED_BOTTLENECK_ANALYSIS.md: Page-fault breakdown
- RANDOM_MIXED_SUMMARY.md: Phase 23 summary
- RING_CACHE_ACTIVATION_GUIDE.md: Ring cache usage
- CURRENT_TASK.md: Updated with Phase 23 results and Phase 24 plan
Next Steps (Phase 24):
- Target: Mid/VM PageArena/HotSpanBox (page-fault reduction 80-100K → 30-40K)
- Tiny SSM optimization deferred (low ROI, ~6K page-faults already optimal)
- Expected improvement: +30-50% for Mid/Large workloads
Generated with Claude Code
Co-Authored-By: Claude <noreply@anthropic.com>
2025-11-17 02:47:58 +09:00
|
|
|
pagefault_telemetry_touch(PF_BUCKET_MID, user4);
|
|
|
|
|
return user4;
|
2025-11-07 01:27:04 +09:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void hak_pool_free(void* ptr, size_t size, uintptr_t site_id) {
|
|
|
|
|
if (!ptr) return;
|
|
|
|
|
hak_pool_init();
|
|
|
|
|
if (!hak_pool_is_poolable(size)) return;
|
|
|
|
|
|
|
|
|
|
if (g_mf2_enabled) { mf2_free(ptr); return; }
|
|
|
|
|
|
|
|
|
|
void* raw = (char*)ptr - HEADER_SIZE;
|
|
|
|
|
AllocHeader* hdr = (AllocHeader*)raw;
|
|
|
|
|
int mid_by_desc = 0; MidPageDesc* d_desc = mid_desc_lookup(ptr);
|
|
|
|
|
if (d_desc) mid_by_desc = 1;
|
|
|
|
|
if (!mid_by_desc && g_hdr_light_enabled < 2) {
|
|
|
|
|
if (hdr->magic != HAKMEM_MAGIC) { MF2_ERROR_LOG("Invalid magic 0x%X in pool_free, expected 0x%X", hdr->magic, HAKMEM_MAGIC); return; }
|
|
|
|
|
if (hdr->method != ALLOC_METHOD_POOL) { MF2_ERROR_LOG("Wrong method %d in pool_free, expected POOL (%d)", hdr->method, ALLOC_METHOD_POOL); return; }
|
|
|
|
|
}
|
|
|
|
|
int class_idx = mid_by_desc ? (int)d_desc->class_idx : hak_pool_get_class_index(size);
|
|
|
|
|
if (class_idx < 0) return;
|
|
|
|
|
PoolBlock* block = (PoolBlock*)raw;
|
|
|
|
|
if (g_pool.tls_free_enabled) {
|
|
|
|
|
int same_thread = 0;
|
|
|
|
|
if (g_hdr_light_enabled >= 1) { MidPageDesc* d = mid_desc_lookup(raw); if (d && d->owner_tid != 0 && d->owner_tid == (uint64_t)(uintptr_t)pthread_self()) { same_thread = 1; } }
|
|
|
|
|
else if (hdr->owner_tid != 0 && hdr->owner_tid == (uintptr_t)(uintptr_t)pthread_self()) { same_thread = 1; }
|
|
|
|
|
if (same_thread) {
|
|
|
|
|
PoolTLSRing* ring = &g_tls_bin[class_idx].ring;
|
|
|
|
|
if (g_tls_ring_enabled && ring->top < POOL_L2_RING_CAP) { ring->items[ring->top++] = block; }
|
|
|
|
|
else { block->next = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = block; g_tls_bin[class_idx].lo_count++; if ((int)g_tls_bin[class_idx].lo_count > g_tls_lo_max) {
|
|
|
|
|
size_t spill = g_tls_bin[class_idx].lo_count / 2; int shard = hak_pool_get_shard_index(site_id);
|
|
|
|
|
while (spill-- && g_tls_bin[class_idx].lo_head) { PoolBlock* b = g_tls_bin[class_idx].lo_head; g_tls_bin[class_idx].lo_head = b->next; g_tls_bin[class_idx].lo_count--; HKM_TIME_START(t_remote_push1); uintptr_t old_head; do { old_head = atomic_load_explicit(&g_pool.remote_head[class_idx][shard], memory_order_acquire); b->next = (PoolBlock*)old_head; } while (!atomic_compare_exchange_weak_explicit(&g_pool.remote_head[class_idx][shard], &old_head, (uintptr_t)b, memory_order_release, memory_order_relaxed)); atomic_fetch_add_explicit(&g_pool.remote_count[class_idx][shard], 1, memory_order_relaxed); HKM_TIME_END(HKM_CAT_POOL_REMOTE_PUSH, t_remote_push1); } set_nonempty_bit(class_idx, shard); } }
|
|
|
|
|
} else {
|
|
|
|
|
if (g_tc_enabled) { uint64_t owner_tid = 0; if (g_hdr_light_enabled < 2) owner_tid = hdr->owner_tid; if (owner_tid == 0) { MidPageDesc* d = mid_desc_lookup(raw); if (d) owner_tid = d->owner_tid; } if (owner_tid != 0) { MidTC* otc = mid_tc_lookup_by_tid(owner_tid); if (otc) { mid_tc_push(otc, class_idx, block); return; } } }
|
|
|
|
|
int shard = hak_pool_get_shard_index(site_id); uintptr_t old_head; HKM_TIME_START(t_remote_push2);
|
|
|
|
|
do { old_head = atomic_load_explicit(&g_pool.remote_head[class_idx][shard], memory_order_acquire); block->next = (PoolBlock*)old_head; } while (!atomic_compare_exchange_weak_explicit(&g_pool.remote_head[class_idx][shard], &old_head, (uintptr_t)block, memory_order_release, memory_order_relaxed));
|
|
|
|
|
atomic_fetch_add_explicit(&g_pool.remote_count[class_idx][shard], 1, memory_order_relaxed); HKM_TIME_END(HKM_CAT_POOL_REMOTE_PUSH, t_remote_push2); set_nonempty_bit(class_idx, shard);
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
int shard_idx2 = hak_pool_get_shard_index(site_id); pthread_mutex_t* lock = &g_pool.freelist_locks[class_idx][shard_idx2].m; pthread_mutex_lock(lock); block->next = g_pool.freelist[class_idx][shard_idx2]; g_pool.freelist[class_idx][shard_idx2] = block; set_nonempty_bit(class_idx, shard_idx2); pthread_mutex_unlock(lock);
|
|
|
|
|
}
|
|
|
|
|
t_pool_rng ^= t_pool_rng << 13; t_pool_rng ^= t_pool_rng >> 17; t_pool_rng ^= t_pool_rng << 5; if ((t_pool_rng & ((1u<<g_count_sample_exp)-1u)) == 0u) g_pool.frees[class_idx]++;
|
|
|
|
|
mid_page_inuse_dec_and_maybe_dn(raw);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
int hak_pool_mid_lookup(void* ptr, size_t* out_size) {
|
|
|
|
|
if (g_mf2_enabled) { MidPage* page = mf2_addr_to_page(ptr); if (page) { int c = (int)page->class_idx; if (c < 0 || c >= POOL_NUM_CLASSES) return 0; size_t sz = g_class_sizes[c]; if (sz == 0) return 0; if (out_size) *out_size = sz; return 1; } }
|
|
|
|
|
MidPageDesc* d = mid_desc_lookup(ptr); if (!d) return 0; int c = (int)d->class_idx; if (c < 0 || c >= POOL_NUM_CLASSES) return 0; size_t sz = g_class_sizes[c]; if (sz == 0) return 0; if (out_size) *out_size = sz; return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
void hak_pool_free_fast(void* ptr, uintptr_t site_id) {
|
|
|
|
|
if (!ptr || !g_pool.initialized) return; if (g_mf2_enabled) { MidPage* page = mf2_addr_to_page(ptr); if (page) { mf2_free(ptr); return; } }
|
|
|
|
|
MidPageDesc* d = mid_desc_lookup(ptr); if (!d) return; size_t sz = g_class_sizes[(int)d->class_idx]; if (sz == 0) return; hak_pool_free(ptr, sz, site_id);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#endif // POOL_API_INC_H
|