Skip to content

Commit

Permalink
ballet/pack: Improve performance test
Browse files Browse the repository at this point in the history
  • Loading branch information
@ptaffet-jump
ptaffet-jump committed Jan 12, 2024
1 parent 0734ef3 commit d23e8fb
Showing 1 changed file with 166 additions and 25 deletions.
191 changes: 166 additions & 25 deletions src/ballet/pack/test_pack.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,6 +174,7 @@ insert( ulong i,
fd_pack_t * pack ) {
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ i ];
slot->payload_sz = payload_sz[ i ];
fd_memcpy( slot->payload, payload_scratch[ i ], payload_sz[ i ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );

Expand Down Expand Up @@ -203,7 +204,6 @@ schedule_validate_microblock( fd_pack_t * pack,

ulong total_rewards = 0UL;


aset_t read_accts = aset_null( );
aset_t write_accts = aset_null( );

Expand Down Expand Up @@ -395,32 +395,169 @@ test_delete( void ) {
FD_TEST( fd_pack_avail_txn_cnt( pack ) == 0UL );
}

void performance_test( void ) {
void performance_test( int extra_bench ) {
ulong i = 0UL;
FD_LOG_NOTICE(( "TEST PERFORMANCE" ));
fd_pack_t * pack = init_all( 10240UL, 1UL, 2UL, &outcome );
make_transaction( i, 800U, 12.0, "ABC", "DEF" );
make_transaction( i+1, 500U, 12.0, "GHJ", "KLMNOP" );
make_transaction( i, 800U, 12.0, "ABC", "DEF" ); /* Total cost 2873 */
make_transaction( i+1, 500U, 12.0, "GHJ", "KLMNOP" ); /* Total cost 2575 */

fd_wksp_t * wksp = fd_wksp_new_anonymous( FD_SHMEM_GIGANTIC_PAGE_SZ, 1UL, 0UL, "test_pack", 0UL );

/* Above 8192, we start running into MAX_WRITE_COST_PER_ACCT.
Obviously, calling end_block fixes that, but it muddies the
performance measurements. */
ulong max_heap_sz = fd_ulong_if( (!!wksp) & extra_bench, 8192UL, 2048UL );
ulong linear_inc = fd_ulong_if( extra_bench, 256UL, 9999999UL );

FD_LOG_NOTICE(( "All columns in units ns/txn except depth and End block (ns/call)" ));
FD_LOG_NOTICE(( "Depth\tPreinsert\tInsert\tEnd block\tSchedule\tSkip fast\tSkip normal 1\tSkip normal 2" ));

#define ITER_CNT 10UL
#define WARMUP 2UL
for( ulong heap_sz=16UL; heap_sz<=max_heap_sz; heap_sz = fd_ulong_min( heap_sz*2UL, heap_sz+linear_inc ) ) {
ulong footprint = fd_pack_footprint( heap_sz, 1UL, 3UL );
void * _mem;
if( FD_LIKELY( wksp ) ) _mem = fd_wksp_alloc_laddr( wksp, fd_pack_align(), footprint, 4UL );
else { FD_TEST( footprint<PACK_SCRATCH_SZ ); _mem = pack_scratch; }

long preinsert = 0L;
long insert = 0L;
long end_block = 0L;
long skip0 = 0L;
long skip1 = 0L;
long skip2 = 0L;
long schedule = 0L;

for( ulong iter=0UL; iter<ITER_CNT; iter++ ) {
fd_pack_t * pack = fd_pack_join( fd_pack_new( _mem, heap_sz, 1UL, 3UL, heap_sz, rng ) );

FD_TEST( fd_pack_avail_txn_cnt( pack )==0UL );

if( FD_LIKELY( iter>=WARMUP ) ) preinsert -= fd_log_wallclock( );
for( ulong j=0UL; j<heap_sz; j++ ) {
memcpy( payload_scratch[j&1]+1UL, &j, sizeof(ulong) );
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ j&1 ];
slot->payload_sz = payload_sz[ j&1 ];
fd_memcpy( slot->payload, payload_scratch[ j&1 ], payload_sz[ j&1 ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );
fd_pack_insert_txn_cancel( pack, slot );
}
if( FD_LIKELY( iter>=WARMUP ) ) preinsert += fd_log_wallclock( );

if( FD_LIKELY( iter>=WARMUP ) ) insert -= fd_log_wallclock( );
for( ulong j=0UL; j<heap_sz; j++ ) {
memcpy( payload_scratch[j&1]+1UL, &j, sizeof(ulong) );
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ j&1 ];
slot->payload_sz = payload_sz[ j&1 ];
fd_memcpy( slot->payload, payload_scratch[ j&1 ], payload_sz[ j&1 ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );

fd_pack_insert_txn_fini( pack, slot );
}
if( FD_LIKELY( iter>=WARMUP ) ) insert += fd_log_wallclock( );

long start = fd_log_wallclock( );
for( ulong j=0UL; j<1024UL; j++ ) {
payload_scratch[j&1][ 1+(j%8) ]++;
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ j&1 ];
fd_memcpy( slot->payload, payload_scratch[ j&1 ], payload_sz[ j&1 ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );
FD_TEST( fd_pack_avail_txn_cnt( pack )==heap_sz );

fd_pack_insert_txn_fini( pack, slot );
}
long end = fd_log_wallclock( );
FD_LOG_NOTICE(( "Inserting when not full: %f ns", ((double)(end-start))/10240.0 ));
start = fd_log_wallclock( );
for( ulong j=0UL; j<10240UL; j++ ) {
fd_pack_schedule_next_microblock( pack, 2000UL, 0.0f, 0UL, outcome.results );
fd_pack_microblock_complete( pack, 0UL );
if( FD_LIKELY( iter>=WARMUP ) ) end_block -= fd_log_wallclock( );
for( ulong j=0UL; j<5UL; j++ ) {
fd_pack_end_block( pack );
}
if( FD_LIKELY( iter>=WARMUP ) ) end_block += fd_log_wallclock( );

if( FD_LIKELY( iter>=WARMUP ) ) skip0 -= fd_log_wallclock( );
for( ulong j=0UL; j<heap_sz/2UL; j++ ) {
/* With a cap of 2000 CUs, nothing fits, but we scan the whole heap
each time to figure that out. */
fd_pack_schedule_next_microblock( pack, 2000UL, 0.0f, 0UL, outcome.results );
fd_pack_microblock_complete( pack, 0UL );
}
if( FD_LIKELY( iter>=WARMUP ) ) skip0 += fd_log_wallclock( );
FD_TEST( fd_pack_avail_txn_cnt( pack )==heap_sz );

fd_pack_end_block( pack );

if( FD_LIKELY( iter>=WARMUP ) ) schedule -= fd_log_wallclock( );
for( ulong j=0UL; j<heap_sz; j++ ) {
/* With a cap of 3000 CUs, we schedule 1 transaction and then
immediately break. */
FD_TEST( 1UL==fd_pack_schedule_next_microblock( pack, 3000UL, 0.0f, 0UL, outcome.results ) );
fd_pack_microblock_complete( pack, 0UL );
}
if( FD_LIKELY( iter>=WARMUP ) ) schedule += fd_log_wallclock( );
FD_TEST( fd_pack_avail_txn_cnt( pack )==0UL );

fd_pack_end_block( pack );

/* Fill the heap back up */
for( ulong j=0UL; j<heap_sz; j++ ) {
memcpy( payload_scratch[j&1]+1UL, &j, sizeof(ulong) );
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ j&1 ];
slot->payload_sz = payload_sz[ j&1 ];
fd_memcpy( slot->payload, payload_scratch[ j&1 ], payload_sz[ j&1 ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );

fd_pack_insert_txn_fini( pack, slot );
}

FD_TEST( fd_pack_avail_txn_cnt( pack )==heap_sz );
if( FD_LIKELY( iter>=WARMUP ) ) skip1 -= fd_log_wallclock( );
for( ulong j=0UL; j<heap_sz/2UL; j++ ) {
/* With a cap of 6000 CUs, we schedule a copy of transaction 1,
scan through all the duplicates of transaction 1 (which
conflict because of accounts), finally find an instance of
transaction 2, schedule it, and then immediately break. */
FD_TEST( 2UL==fd_pack_schedule_next_microblock( pack, 6000UL, 0.0f, 0UL, outcome.results ) );
fd_pack_microblock_complete( pack, 0UL );
}
if( FD_LIKELY( iter>=WARMUP ) ) skip1 += fd_log_wallclock( );
FD_TEST( fd_pack_avail_txn_cnt( pack )==0UL );

fd_pack_end_block( pack );

/* Fill the heap back up */
for( ulong j=0UL; j<heap_sz; j++ ) {
memcpy( payload_scratch[j&1]+1UL, &j, sizeof(ulong) );
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ j&1 ];
slot->payload_sz = payload_sz[ j&1 ];
fd_memcpy( slot->payload, payload_scratch[ j&1 ], payload_sz[ j&1 ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );

fd_pack_insert_txn_fini( pack, slot );
}

FD_TEST( fd_pack_avail_txn_cnt( pack )==heap_sz );
if( FD_LIKELY( iter>=WARMUP ) ) skip2 -= fd_log_wallclock( );
for( ulong j=0UL; j<heap_sz/2UL; j++ ) {
/* With a huge CU cap, we schedule a copy of transaction 1,
scan through all the duplicates of transaction 1 (which
conflict because of accounts), finally find an instance of
transaction 2, schedule it, then continue skipping through
all the copies of transaction 2. */
FD_TEST( 2UL==fd_pack_schedule_next_microblock( pack, 5000000UL, 0.0f, 0UL, outcome.results ) );
fd_pack_microblock_complete( pack, 0UL );
}
if( FD_LIKELY( iter>=WARMUP ) ) skip2 += fd_log_wallclock( );
FD_TEST( fd_pack_avail_txn_cnt( pack )==0UL );
}
double denominator = (double)(heap_sz*(ITER_CNT-WARMUP));
FD_LOG_NOTICE(( "%5lu\t%9.3f\t%6.3f\t%9.0f\t%8.3f\t%9.3f\t%13.3f\t%13.3f", heap_sz,
(double)(preinsert )/denominator,
(double)(insert-preinsert )/denominator,
(double)(end_block )/(double)5.0,
(double)(schedule )/denominator,
(double)(skip0 )/(denominator*(double)(heap_sz )/2.0),
(double)(skip1-schedule )/(denominator*(double)(heap_sz+2UL)/8.0),
(double)(skip2-schedule )/(denominator*(double)(heap_sz+2UL)/4.0) ));


if( FD_LIKELY( wksp ) ) fd_wksp_free_laddr( _mem );
}
end = fd_log_wallclock( );
FD_LOG_NOTICE(( "Scheduling: %f ns", ((double)(end-start))/10240.0 ));
#undef WARMUP
#undef ITER_CNT
}


Expand All @@ -432,20 +569,22 @@ void heap_overflow_test( void ) {
make_transaction( j, 800U, 4.0, "ABC", "DEF" );
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ j ];
slot->payload_sz = payload_sz[ j ];
fd_memcpy( slot->payload, payload_scratch[ j ], payload_sz[ j ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );

fd_pack_insert_txn_fini( pack, slot );
}
FD_TEST( fd_pack_avail_txn_cnt( pack )==1024UL );

/* Now insert higher-paying transactions. They should mostly take the place
of the low-paying transactions */
/* Now insert higher-paying transactions. They should take the
place of the low-paying transactions */
ulong r_hi = make_transaction( 1UL, 500U, 10.0, "GHJ", "KLMNOP" );
for( ulong j=0UL; j<1024UL; j++ ) {
payload_scratch[1][ 1+(j%8) ]++;
fd_txn_p_t * slot = fd_pack_insert_txn_init( pack );
fd_txn_t * txn = (fd_txn_t*) txn_scratch[ 1UL ];
slot->payload_sz = payload_sz[ 1UL ];
fd_memcpy( slot->payload, payload_scratch[ 1UL ], payload_sz[ 1UL ] );
fd_memcpy( TXN(slot), txn, fd_txn_footprint( txn->instr_cnt, txn->addr_table_lookup_cnt ) );

Expand Down Expand Up @@ -651,16 +790,18 @@ main( int argc,
rng = fd_rng_join( fd_rng_new( _rng, 0U, 0UL ) );
fd_metrics_register( (ulong *)fd_metrics_new( metrics_scratch, 0UL, 0UL ) );

int extra_benchmark = fd_env_strip_cmdline_contains( &argc, &argv, "--extra-bench" );

test0();
test1();
test2();
test_vote();
performance_test();
heap_overflow_test();
test_delete();
test_gap();
test_limits();
test_reject_writes_to_sysvars();
performance_test( extra_benchmark );

fd_rng_delete( fd_rng_leave( rng ) );

Expand Down

0 comments on commit d23e8fb

Please sign in to comment.