Skip to content

Benchmarks: Sweep accdb Cache Footprint and Min Reserved #11437

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MilanSuri wants to merge 4 commits into
base: main
Choose a base branch
from
+95 −18
Open

Benchmarks: Sweep accdb Cache Footprint and Min Reserved #11437

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f578549
Bench: add accdb cache footprint sweep
MilanSuri Sep 12, 2026
bc3d7ce
Bench: sweep accdb cache sizing and retain stable rng throughout the …
MilanSuri Sep 12, 2026
fea9adb
Formatting: fixed spacing and comment
MilanSuri Sep 12, 2026
905d739
Formatting: fixed coment and formatting of bitshift
MilanSuri Sep 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
113 changes: 95 additions & 18 deletions src/flamenco/accdb/bench_accdb.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,31 +82,41 @@ static uchar data_buf[ BENCH_MAX_DATA_SZ ];
#define BENCH_CACHE_MIN_RESERVED (640UL)

static fd_accdb_t *
bench_setup( int * out_fd,
ulong max_accounts,
ulong max_live_slots,
ulong max_account_writes_per_slot,
ulong partition_cnt,
ulong partition_sz ) {
bench_setup( int * out_fd,
void ** out_shmem_mem,
void ** out_accdb_mem,
ulong max_accounts,
ulong max_live_slots,
ulong max_account_writes_per_slot,
ulong partition_cnt,
ulong partition_sz,
ulong cache_footprint,
ulong cache_min_reserved) {
int fd = memfd_create( "accdb_bench", 0 );
if( FD_UNLIKELY( fd<0 ) ) FD_LOG_ERR(( "memfd_create failed" ));
*out_fd = fd;

ulong cache_fp = BENCH_CACHE_FOOTPRINT;
ulong shmem_fp = fd_accdb_shmem_footprint( max_accounts, max_live_slots, max_account_writes_per_slot, partition_cnt, cache_fp, BENCH_CACHE_MIN_RESERVED, 1UL, 0UL );
ulong cache_fp = cache_footprint;
ulong shmem_fp = fd_accdb_shmem_footprint( max_accounts, max_live_slots, max_account_writes_per_slot, partition_cnt, cache_fp, cache_min_reserved, 1UL, 0UL );
FD_TEST( shmem_fp );

void * shmem_mem = aligned_alloc( fd_accdb_shmem_align(), shmem_fp );
FD_TEST( shmem_mem );
*out_shmem_mem = shmem_mem;

fd_accdb_shmem_t * shmem = fd_accdb_shmem_join(
fd_accdb_shmem_new( shmem_mem, max_accounts, max_live_slots,
max_account_writes_per_slot, partition_cnt,
partition_sz, cache_fp, BENCH_CACHE_MIN_RESERVED, 0, 42UL, 1UL, 0UL ) );
partition_sz, cache_fp, cache_min_reserved, 0, 42UL, 1UL, 0UL ) );
FD_TEST( shmem );

ulong accdb_fp = fd_accdb_footprint( max_live_slots );
FD_TEST( accdb_fp );

void * accdb_mem = aligned_alloc( fd_accdb_align(), accdb_fp );
FD_TEST( accdb_mem );
*out_accdb_mem = accdb_mem;

fd_accdb_t * accdb = fd_accdb_join( fd_accdb_new( accdb_mem, shmem, fd, 0UL, NULL ) );
FD_TEST( accdb );
return accdb;
Expand Down Expand Up @@ -170,14 +180,20 @@ static void
bench_write( ulong account_cnt,
fd_rng_t * rng ) {
int fd;
void * shmem_mem;
void * accdb_mem;
ulong partition_sz = 1UL<<30UL;
ulong partition_cnt = 1024UL;
fd_accdb_t * accdb = bench_setup( &fd,
&shmem_mem,
&accdb_mem,
account_cnt + 1024UL,
64UL,
(uint)account_cnt + 1024U,
partition_cnt,
partition_sz );
partition_sz,
BENCH_CACHE_FOOTPRINT,
BENCH_CACHE_MIN_RESERVED);

fd_accdb_fork_id_t root = fd_accdb_attach_child( accdb, (fd_accdb_fork_id_t){ .val = USHORT_MAX } );
fd_accdb_fork_id_t fork = fd_accdb_attach_child( accdb, root );
Expand Down Expand Up @@ -206,6 +222,8 @@ bench_write( ulong account_cnt,
(double)total_bytes / (double)(1UL<<20UL) / secs,
(double)dt / (double)account_cnt ));

free( accdb_mem );
free( shmem_mem );
close( fd );
}

Expand All @@ -217,14 +235,20 @@ static void
bench_read( ulong account_cnt,
fd_rng_t * rng ) {
int fd;
void * shmem_mem;
void * accdb_mem;
ulong partition_sz = 1UL<<30UL;
ulong partition_cnt = 1024UL;
fd_accdb_t * accdb = bench_setup( &fd,
&shmem_mem,
&accdb_mem,
account_cnt + 1024UL,
64UL,
(uint)account_cnt + 1024U,
partition_cnt,
partition_sz );
partition_sz,
BENCH_CACHE_FOOTPRINT,
BENCH_CACHE_MIN_RESERVED);

fd_accdb_fork_id_t root = fd_accdb_attach_child( accdb, (fd_accdb_fork_id_t){ .val = USHORT_MAX } );
fd_accdb_fork_id_t fork = fd_accdb_attach_child( accdb, root );
Expand Down Expand Up @@ -269,6 +293,8 @@ bench_read( ulong account_cnt,
(double)total_bytes / (double)(1UL<<20UL) / secs,
(double)dt / (double)account_cnt ));

free( accdb_mem );
free( shmem_mem );
close( fd );
}

Expand All @@ -292,18 +318,26 @@ static void
bench_replay( ulong slot_cnt,
ulong writes_per_slot,
ulong reads_per_slot,
fd_rng_t * rng ) {
fd_rng_t * rng,
ulong cache_footprint,
ulong cache_min_reserve ) {
int fd;
void * shmem_mem;
void * accdb_mem;
ulong total_accounts = slot_cnt * writes_per_slot + 1024UL;
ulong partition_sz = 1UL<<30UL;
ulong partition_cnt = 1024UL;
ulong max_live = 64UL;
fd_accdb_t * accdb = bench_setup( &fd,
&shmem_mem,
&accdb_mem,
total_accounts,
max_live,
writes_per_slot + 16UL,
partition_cnt,
partition_sz );
partition_sz,
cache_footprint,
cache_min_reserve );

fd_accdb_fork_id_t root = fd_accdb_attach_child( accdb, (fd_accdb_fork_id_t){ .val = USHORT_MAX } );

Expand All @@ -312,7 +346,7 @@ bench_replay( ulong slot_cnt,
/* We re-use a pool of pubkeys so that rooting has old versions to
tombstone (realistic for hot accounts like token program,
system program, fee payers, etc.). */
ulong pubkey_pool_sz = writes_per_slot * 4UL;
ulong pubkey_pool_sz = writes_per_slot * slot_cnt;

@MilanSuri MilanSuri Sep 12, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah I found this issue as well, I changed the pubkey_pool_sz from writes_per_slot * 4UL to being writes_per_slot * slot_cnt because I needed a way of increasing cache pressure for each sweep.

Is there a way to get the true cache hit rate in the Firedancer codebase or would it be better to just rename it to like found rate?


ulong total_reads = 0UL;
ulong total_read_hits = 0UL;
Expand Down Expand Up @@ -405,6 +439,8 @@ bench_replay( ulong slot_cnt,
(double)m->disk_used_bytes / (double)(1UL<<20UL),
(double)m->disk_allocated_bytes / (double)(1UL<<20UL) ));

free( accdb_mem );
free( shmem_mem );
close( fd );
}

Expand All @@ -420,14 +456,20 @@ bench_mixed( ulong base_cnt,
uint read_pct,
fd_rng_t * rng ) {
int fd;
void * shmem_mem;
void * accdb_mem;
ulong partition_sz = 1UL<<30UL;
ulong partition_cnt = 1024UL;
fd_accdb_t * accdb = bench_setup( &fd,
&shmem_mem,
&accdb_mem,
base_cnt + op_cnt + 1024UL,
64UL,
(uint)(base_cnt + op_cnt) + 1024U,
partition_cnt,
partition_sz );
partition_sz,
BENCH_CACHE_FOOTPRINT,
BENCH_CACHE_MIN_RESERVED);

fd_accdb_fork_id_t root = fd_accdb_attach_child( accdb, (fd_accdb_fork_id_t){ .val = USHORT_MAX } );
fd_accdb_fork_id_t fork = fd_accdb_attach_child( accdb, root );
Expand Down Expand Up @@ -476,6 +518,8 @@ bench_mixed( ulong base_cnt,
(double)dt / (double)op_cnt,
(uint)read_pct ));

free( accdb_mem );
free( shmem_mem );
close( fd );
}

Expand All @@ -492,7 +536,7 @@ main( int argc,
ulong reads_per_slot = fd_env_strip_cmdline_ulong( &argc, &argv, "--reads-per-slot", NULL, 4800UL );
ulong mixed_ops = fd_env_strip_cmdline_ulong( &argc, &argv, "--mixed-ops", NULL, 200000UL );
uint read_pct = fd_env_strip_cmdline_uint ( &argc, &argv, "--read-pct", NULL, 80U );
uint seed = fd_env_strip_cmdline_uint ( &argc, &argv, "--seed", NULL, 42U );
uint seed = fd_env_strip_cmdline_uint ( &argc, &argv, "--seed", NULL, 42U );

FD_LOG_NOTICE(( "accdb benchmark (accounts=%lu slots=%lu wpslot=%lu "
"rpslot=%lu mixed_ops=%lu read_pct=%u seed=%u)",
Expand All @@ -510,14 +554,47 @@ main( int argc,
bench_read( account_cnt, rng );

FD_LOG_NOTICE(( "--- replay simulation ---" ));
bench_replay( slot_cnt, writes_per_slot, reads_per_slot, rng );
bench_replay( slot_cnt, writes_per_slot, reads_per_slot, rng, BENCH_CACHE_FOOTPRINT, BENCH_CACHE_MIN_RESERVED);

FD_LOG_NOTICE(( "--- mixed read/write (%u%% reads) ---", read_pct ));
bench_mixed( account_cnt, mixed_ops, read_pct, rng );

FD_LOG_NOTICE(( "--- cache sweep ---" ));
ulong const cache_sweep_gib[] = { 7UL, 8UL, 10UL, 12UL, 16UL };
int length = sizeof( cache_sweep_gib ) / sizeof( cache_sweep_gib[0] );

for( int i = 0; i < length; i++ ) {
fd_rng_t _sweep_rng[1];
fd_rng_t * sweep_rng = fd_rng_join ( fd_rng_new ( _sweep_rng, seed, 0UL ) );

FD_TEST( sweep_rng );

FD_LOG_NOTICE(( "--- cache sweep @ %lu GiB ---", cache_sweep_gib[i] ));
bench_replay( 1000, writes_per_slot, reads_per_slot, sweep_rng, cache_sweep_gib[i] << 30U, BENCH_CACHE_MIN_RESERVED );

fd_rng_delete( fd_rng_leave ( sweep_rng ));
}

FD_LOG_NOTICE(( "--- cache min reserved sweep ---" ));
ulong const min_reserved_sweep[] = { 160UL, 320UL, 480UL, 640UL, 800UL };
length = sizeof( min_reserved_sweep ) / sizeof( min_reserved_sweep[0] );

for ( int i = 0; i < length; i++ ) {
fd_rng_t _sweep_rng[1];
fd_rng_t * sweep_rng = fd_rng_join( fd_rng_new ( _sweep_rng, seed, 0UL ));

FD_TEST( sweep_rng );

FD_LOG_NOTICE(( "--- cache min reserved sweep @ %lu ---", min_reserved_sweep[i] ));
/* Using 10 GiB based on cache footprint sweep and to support min_reserved = 800 */
bench_replay( 1000, writes_per_slot, reads_per_slot, sweep_rng, 10UL << 30UL, min_reserved_sweep[i] );

fd_rng_delete( fd_rng_leave ( sweep_rng ));
}

fd_rng_delete( fd_rng_leave( rng ) );

FD_LOG_NOTICE(( "pass" ));
fd_halt();
return 0;
}
}