bench: firedancer integration

Author: mmcgee-jump

book/api/metrics-generated.md

@@ -1106,3 +1106,13 @@
 | <span class="metrics-name">snapwr_&#8203;vinyl_&#8203;bytes_&#8203;written</span> | gauge | Number of bytes written so far to the vinyl snapshot file. Might decrease if snapshot creation is aborted and restarted |
 
 </div>
+
+## Benchs Tile
+
+<div class="metrics">
+
+| Metric | Type | Description |
+|--------|------|-------------|
+| <span class="metrics-name">benchs_&#8203;transactions_&#8203;sent</span> | counter | Number of benchmark packets sent |
+
+</div>

src/app/fddev/commands/bench.c

@@ -20,7 +20,7 @@ agave_thread_main( void * _args ) {
 void
 fddev_bench_cmd_fn( args_t *   args,
                     config_t * config ) {
-  bench_cmd_fn( args, config );
+  bench_cmd_fn( args, config, 0 );
 
   pthread_t agave;
   pthread_create( &agave, NULL, agave_thread_main, (void *)config );

src/app/firedancer-dev/commands/backtest.c

@@ -261,7 +261,7 @@ backtest_topo( config_t * config ) {
 
   /* banks_obj shared by replay and exec tiles */
   fd_topob_wksp( topo, "banks" );
-  fd_topo_obj_t * banks_obj = setup_topo_banks( topo, "banks", config->firedancer.runtime.max_live_slots, config->firedancer.runtime.max_fork_width );
+  fd_topo_obj_t * banks_obj = setup_topo_banks( topo, "banks", config->firedancer.runtime.max_live_slots, config->firedancer.runtime.max_fork_width, 0 );
   fd_topob_tile_uses( topo, replay_tile, banks_obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
   FOR(exec_tile_cnt) fd_topob_tile_uses( topo, &topo->tiles[ fd_topo_find_tile( topo, "exec", i ) ], banks_obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
   FD_TEST( fd_pod_insertf_ulong( topo->props, banks_obj->id, "banks" ) );

src/app/firedancer-dev/commands/bench.c

@@ -3,10 +3,16 @@
 
 #include <unistd.h>
 
+static void
+bench_cmd_topo( config_t * config ) {
+  config->development.sandbox  = 0;
+  config->development.no_clone = 1;
+}
+
 void
 firedancer_dev_bench_cmd_fn( args_t *   args,
                              config_t * config ) {
-  bench_cmd_fn( args, config );
+  bench_cmd_fn( args, config, 1 );
 
   /* Sleep parent thread forever, Ctrl+C will terminate. */
   for(;;) pause();
@@ -17,6 +23,7 @@ action_t fd_action_bench = {
   .args             = bench_cmd_args,
   .fn               = firedancer_dev_bench_cmd_fn,
   .perm             = dev_cmd_perm,
+  .topo             = bench_cmd_topo,
   .is_local_cluster = 1,
   .description      = "Test validator TPS benchmark"
 };

src/app/firedancer-dev/commands/sim.c

@@ -126,7 +126,7 @@ sim_topo( config_t * config ) {
       config->firedancer.runtime.max_live_slots,
       fd_ulong_pow2_up( FD_PACK_MAX_TXNCACHE_TXN_PER_SLOT ) );
   fd_topo_obj_t * poh_slot_obj = fd_topob_obj( topo, "fseq", "poh_slot" );
-  fd_topo_obj_t * banks_obj = setup_topo_banks( topo, "banks", config->firedancer.runtime.max_live_slots, config->firedancer.runtime.max_fork_width );
+  fd_topo_obj_t * banks_obj = setup_topo_banks( topo, "banks", config->firedancer.runtime.max_live_slots, config->firedancer.runtime.max_fork_width, 0 );
 
   FD_TEST( fd_pod_insertf_ulong( topo->props, poh_shred_obj->id, "poh_shred" ) );
   FD_TEST( fd_pod_insertf_ulong( topo->props, root_slot_obj->id, "root_slot" ) );

src/app/firedancer/callbacks.c

@@ -26,7 +26,9 @@ banks_align( fd_topo_t const *     topo FD_FN_UNUSED,
 static void
 banks_new( fd_topo_t const *     topo,
            fd_topo_obj_t const * obj ) {
-  FD_TEST( fd_banks_new( fd_topo_obj_laddr( topo, obj->id ), VAL("max_live_slots"), VAL("max_fork_width") ) );
+  int larger_max_cost_per_block = fd_pod_queryf_int( topo->props, 0, "obj.%lu.larger_max_cost_per_block", obj->id );
+  ulong seed = fd_pod_queryf_ulong( topo->props, 0UL, "obj.%lu.seed", obj->id );
+  FD_TEST( fd_banks_new( fd_topo_obj_laddr( topo, obj->id ), VAL("max_live_slots"), VAL("max_fork_width"), larger_max_cost_per_block, seed ) );
 }
 
 fd_topo_obj_callbacks_t fd_obj_cb_banks = {

src/app/firedancer/config/bench-zen3-32core.toml

@@ -0,0 +1,34 @@
+# A suggested configuration for benchmarking on a Zen3 CPU like AMD EPYC
+# 7513 or similar.  The configuration expects 64 logical cores, and 32
+# physical cores, numbered 0 and 1 for thread 0 and thread 1 on core 0,
+# 2 and 3 for thread 0 and thread 1 on core 1, and so on.  In
+# production, the configuration would not be correct as cores do not
+# need to be dedicated to generating and sending transactions for
+# benchmarking.
+[layout]
+  affinity = "14-63,f1"
+  net_tile_count = 1
+  quic_tile_count = 1
+  resolv_tile_count = 1
+  verify_tile_count = 20
+  gossvf_tile_count = 1
+  bank_tile_count = 4
+  exec_tile_count = 1
+  shred_tile_count = 1
+  sign_tile_count = 2
+
+[development.genesis]
+  fund_initial_accounts = 32768
+
+[development.bench]
+  benchg_tile_count = 12
+  benchs_tile_count = 2
+  affinity = "f1,0-13"
+  larger_max_cost_per_block = true
+  larger_shred_limits_per_block = true
+
+[tiles.shred]
+  max_pending_shred_sets = 16384
+
+[tiles.pack]
+  schedule_strategy = "perf"

src/app/firedancer/topology.c

@@ -59,10 +59,15 @@ fd_topo_obj_t *
 setup_topo_banks( fd_topo_t *  topo,
                   char const * wksp_name,
                   ulong        max_live_slots,
-                  ulong        max_fork_width ) {
+                  ulong        max_fork_width,
+                  int          larger_max_cost_per_block ) {
   fd_topo_obj_t * obj = fd_topob_obj( topo, "banks", wksp_name );
   FD_TEST( fd_pod_insertf_ulong( topo->props, max_live_slots, "obj.%lu.max_live_slots", obj->id ) );
   FD_TEST( fd_pod_insertf_ulong( topo->props, max_fork_width, "obj.%lu.max_fork_width", obj->id ) );
+  FD_TEST( fd_pod_insertf_int( topo->props, larger_max_cost_per_block, "obj.%lu.larger_max_cost_per_block", obj->id ) );
+  ulong seed;
+  FD_TEST( fd_rng_secure( &seed, sizeof( ulong ) ) );
+  FD_TEST( fd_pod_insertf_ulong( topo->props, seed, "obj.%lu.seed", obj->id ) );
   return obj;
 }
 
@@ -809,7 +814,7 @@ fd_topo_initialize( config_t * config ) {
   FOR(bank_tile_cnt)   fd_topob_tile_uses( topo, &topo->tiles[ fd_topo_find_tile( topo, "bank",   i   ) ], funk_obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
   FOR(resolv_tile_cnt) fd_topob_tile_uses( topo, &topo->tiles[ fd_topo_find_tile( topo, "resolv", i   ) ], funk_obj, FD_SHMEM_JOIN_MODE_READ_ONLY  );
 
-  fd_topo_obj_t * banks_obj = setup_topo_banks( topo, "banks", config->firedancer.runtime.max_live_slots, config->firedancer.runtime.max_fork_width );
+  fd_topo_obj_t * banks_obj = setup_topo_banks( topo, "banks", config->firedancer.runtime.max_live_slots, config->firedancer.runtime.max_fork_width, config->development.bench.larger_max_cost_per_block );
   /**/                 fd_topob_tile_uses( topo, &topo->tiles[ fd_topo_find_tile( topo, "replay", 0UL ) ], banks_obj, FD_SHMEM_JOIN_MODE_READ_WRITE ); /* TODO: Should be readonly? */
   FOR(exec_tile_cnt)   fd_topob_tile_uses( topo, &topo->tiles[ fd_topo_find_tile( topo, "exec",   i   ) ], banks_obj, FD_SHMEM_JOIN_MODE_READ_WRITE ); /* TODO: Should be readonly? */
   FOR(bank_tile_cnt)   fd_topob_tile_uses( topo, &topo->tiles[ fd_topo_find_tile( topo, "bank",   i   ) ], banks_obj, FD_SHMEM_JOIN_MODE_READ_WRITE );
@@ -1091,6 +1096,8 @@ fd_topo_configure_tile( fd_topo_tile_t * tile,
 
     tile->replay.expected_shred_version = config->consensus.expected_shred_version;
 
+    tile->replay.larger_max_cost_per_block = config->development.bench.larger_max_cost_per_block;
+
     /* not specified by [tiles.replay] */
 
     strncpy( tile->replay.identity_key_path, config->paths.identity_key, sizeof(tile->replay.identity_key_path) );

src/app/firedancer/topology.h

@@ -20,7 +20,8 @@ fd_topo_obj_t *
 setup_topo_banks( fd_topo_t *  topo,
                   char const * wksp_name,
                   ulong        max_live_slots,
-                  ulong        max_fork_width );
+                  ulong        max_fork_width,
+                  int          larger_max_cost_per_block );
 
 fd_topo_obj_t *
 setup_topo_funk( fd_topo_t *  topo,

src/app/shared/commands/watch/watch.c

@@ -173,6 +173,8 @@ total_regime( ulong const * metrics ) {
   return sum;
 }
 
+static ulong tps_sent_samples_idx = 0UL;
+static ulong tps_sent_samples[ 200UL ];
 static ulong sps_samples_idx = 0UL;
 static ulong sps_samples[ 200UL ];
 static ulong tps_samples_idx = 0UL;
@@ -213,6 +215,45 @@ static ulong snapshot_acc_samples[ 100UL ];
     fmt_countf( fd_alloca_check( 1UL, 64UL ), 64UL, count ); \
   }))
 
+static int
+write_bench( config_t const * config,
+             ulong const *    cur_tile,
+             ulong const *    prev_tile ) {
+  if( FD_UNLIKELY( fd_topo_find_tile( &config->topo, "benchs", 0UL )==ULONG_MAX ) ) return 0;
+
+  ulong tps_sum = 0UL;
+  ulong num_tps_samples = fd_ulong_min( tps_sent_samples_idx, sizeof(tps_sent_samples)/sizeof(tps_sent_samples[0]));
+  for( ulong i=0UL; i<num_tps_samples; i++ ) tps_sum += tps_sent_samples[ i ];
+  char * tps_str = COUNTF( 100.0*(double)tps_sum/(double)num_tps_samples );
+
+  PRINT( "🌶  \033[1m\033[92mBENCH.......\033[0m\033[22m \033[1mGENERATED TPS\033[22m %s \033[1mBENCHG BUSY\033[22m", tps_str );
+  for( ulong i=0UL; i<config->topo.tile_cnt; i++ ) {
+    if( FD_LIKELY( strcmp( config->topo.tiles[ i ].name, "benchg" ) ) ) continue;
+
+    ulong total_ticks = total_regime( &cur_tile[ i*FD_METRICS_TOTAL_SZ ] )-total_regime( &prev_tile[ i*FD_METRICS_TOTAL_SZ ] );
+    double backp_pct = 100.0*(double)( diff_tile( config, "benchg", prev_tile, cur_tile, MIDX( COUNTER, TILE, REGIME_DURATION_NANOS_BACKPRESSURE_PREFRAG ) ) )/(double)total_ticks;
+    double idle_pct = 100.0*(double)( diff_tile( config, "benchg", prev_tile, cur_tile, MIDX( COUNTER, TILE, REGIME_DURATION_NANOS_CAUGHT_UP_POSTFRAG ) ) )/(double)total_ticks;
+    double busy_pct = 100.0 - idle_pct - backp_pct;
+
+    PRINT( " %.1f %%", busy_pct );
+  }
+
+  PRINT( " \033[1mBENCHS BUSY\033[22m" );
+  for( ulong i=0UL; i<config->topo.tile_cnt; i++ ) {
+    if( FD_LIKELY( strcmp( config->topo.tiles[ i ].name, "benchs" ) ) ) continue;
+
+    ulong total_ticks = total_regime( &cur_tile[ i*FD_METRICS_TOTAL_SZ ] )-total_regime( &prev_tile[ i*FD_METRICS_TOTAL_SZ ] );
+    double backp_pct = 100.0*(double)( diff_tile( config, "benchs", prev_tile, cur_tile, MIDX( COUNTER, TILE, REGIME_DURATION_NANOS_BACKPRESSURE_PREFRAG ) ) )/(double)total_ticks;
+    double idle_pct = 100.0*(double)( diff_tile( config, "benchs", prev_tile, cur_tile, MIDX( COUNTER, TILE, REGIME_DURATION_NANOS_CAUGHT_UP_POSTFRAG ) ) )/(double)total_ticks;
+    double busy_pct = 100.0 - idle_pct - backp_pct;
+
+    PRINT( " %.1f %%", busy_pct );
+  }
+
+  PRINT( "\033[K\n" );
+  return 1;
+}
+
 static void
 write_backtest( config_t const * config,
                 ulong const *    cur_tile ) {
@@ -397,7 +438,7 @@ write_replay( config_t const * config,
   for( ulong i=0UL; i<num_tps_samples; i++ ) tps_sum += tps_samples[ i ];
   char * tps_str = COUNTF( 100.0*(double)tps_sum/(double)num_tps_samples );
 
-  PRINT( "💥 \033[1m\033[35mREPLAY......\033[0m\033[22m \033[1mSLOT\033[22m %lu (%ld) \033[1mTPS\033[22m %s \033[1mSPS\033[22m %s \033[1mLEADER IN\033[22m %s \033[1mROOT DIST\033[22m %lu \033[1mBANKS\033[22m %lu\033[K\n",
+  PRINT( "💥 \033[1m\033[35mREPLAY......\033[0m\033[22m \033[1mSLOT\033[22m %lu (%02ld) \033[1mTPS\033[22m %s \033[1mSPS\033[22m %s \033[1mLEADER IN\033[22m %s \033[1mROOT DIST\033[22m %lu \033[1mBANKS\033[22m %lu\033[K\n",
     reset_slot,
     (long)reset_slot-(long)turbine_slot,
     tps_str,
@@ -433,6 +474,8 @@ write_summary( config_t const * config,
 
   lines_printed = 1UL;
 
+  if( FD_UNLIKELY( write_bench( config, cur_tile, prev_tile ) ) ) lines_printed++;
+
   ulong backt_idx = fd_topo_find_tile( &config->topo, "backt", 0UL );
   if( FD_UNLIKELY( backt_idx!=ULONG_MAX ) ) {
     lines_printed++;
@@ -527,6 +570,9 @@ run( config_t const * config,
       snap_tiles( &config->topo, tiles+last_snap*tile_cnt*FD_METRICS_TOTAL_SZ );
       snap_links( &config->topo, links+last_snap*(cons_cnt*8UL*FD_METRICS_ALL_LINK_IN_TOTAL) );
 
+      tps_sent_samples[ tps_sent_samples_idx%(sizeof(tps_sent_samples)/sizeof(tps_sent_samples[0])) ] = (ulong)diff_tile( config, "benchs", tiles+(1UL-last_snap)*tile_cnt*FD_METRICS_TOTAL_SZ, tiles+last_snap*tile_cnt*FD_METRICS_TOTAL_SZ, MIDX( COUNTER, BENCHS, TRANSACTIONS_SENT ) );
+      tps_sent_samples_idx++;
+
       sps_samples[ sps_samples_idx%(sizeof(sps_samples)/sizeof(sps_samples[0])) ] = (ulong)diff_tile( config, "replay", tiles+(1UL-last_snap)*tile_cnt*FD_METRICS_TOTAL_SZ, tiles+last_snap*tile_cnt*FD_METRICS_TOTAL_SZ, MIDX( COUNTER, REPLAY, SLOTS_TOTAL ) );
       sps_samples_idx++;
       tps_samples[ tps_samples_idx%(sizeof(tps_samples)/sizeof(tps_samples[0])) ] = (ulong)diff_tile( config, "replay", tiles+(1UL-last_snap)*tile_cnt*FD_METRICS_TOTAL_SZ, tiles+last_snap*tile_cnt*FD_METRICS_TOTAL_SZ, MIDX( COUNTER, REPLAY, TRANSACTIONS_TOTAL ) );