Skip to content

repair: coarse peer selection #6664

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
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

repair: coarse peer selection #6664

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
392ecd7
repair tool: fix
emwang-jump Oct 2, 2025
719a33d
repair: asyncify signs for init requests and pongs
emwang-jump Oct 2, 2025
a9f85e6
repair: coarse peer selection
emwang-jump Oct 7, 2025
a2aebe8
metric conversion
emwang-jump Oct 10, 2025
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
41 changes: 22 additions & 19 deletions src/app/firedancer-dev/commands/repair.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,8 +151,6 @@ repair_topo( config_t * config ) {

FOR(shred_tile_cnt) fd_topob_link( topo, "repair_shred", "shred_out", pending_fec_shreds_depth, sizeof(fd_ed25519_sig_t), 1UL );

/**/ fd_topob_link( topo, "ping_sign", "repair_sign", 128UL, FD_REPAIR_MAX_PREIMAGE_SZ, 1UL );
/**/ fd_topob_link( topo, "sign_ping", "sign_repair", 128UL, sizeof(fd_ed25519_sig_t), 1UL );
FOR(sign_tile_cnt-1) fd_topob_link( topo, "repair_sign", "repair_sign", 128UL, FD_REPAIR_MAX_PREIMAGE_SZ, 1UL );
FOR(sign_tile_cnt-1) fd_topob_link( topo, "sign_repair", "sign_repair", 128UL, sizeof(fd_ed25519_sig_t), 1UL );

Expand Down Expand Up @@ -246,7 +244,6 @@ repair_topo( config_t * config ) {
/**/ fd_topob_tile_in( topo, "shred", i, "metric_in", "sign_shred", i, FD_TOPOB_UNRELIABLE, FD_TOPOB_UNPOLLED );
/**/ fd_topob_tile_out( topo, "sign", 0UL, "sign_shred", i );
}

FOR(gossvf_tile_cnt) fd_topob_tile_in ( topo, "gossvf", i, "metric_in", "replay_stake", 0UL, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );

/**/ fd_topob_tile_in ( topo, "gossip", 0UL, "metric_in", "replay_stake", 0UL, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
Expand All @@ -256,19 +253,13 @@ repair_topo( config_t * config ) {
/**/ fd_topob_tile_in( topo, "repair", 0UL, "metric_in", "replay_stake", 0UL, FD_TOPOB_UNRELIABLE, FD_TOPOB_POLLED );
fd_topob_tile_in( topo, "repair", 0UL, "metric_in", "snap_out", 0UL, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
FOR(shred_tile_cnt) fd_topob_tile_in( topo, "repair", 0UL, "metric_in", "shred_out", i, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
FOR(shred_tile_cnt) fd_topob_tile_out( topo, "repair", 0UL, "repair_shred", i );
FOR(sign_tile_cnt-1) fd_topob_tile_out( topo, "repair", 0UL, "repair_sign", i );
FOR(sign_tile_cnt-1) fd_topob_tile_in ( topo, "sign", i+1, "metric_in", "repair_sign", i, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
FOR(sign_tile_cnt-1) fd_topob_tile_out( topo, "sign", i+1, "sign_repair", i );
FOR(sign_tile_cnt-1) fd_topob_tile_in ( topo, "repair", 0UL, "metric_in", "sign_repair", i, FD_TOPOB_UNRELIABLE, FD_TOPOB_POLLED );

/**/ fd_topob_tile_in( topo, "sign", 0UL, "metric_in", "ping_sign", 0UL, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
/**/ fd_topob_tile_out( topo, "repair", 0UL, "ping_sign", 0UL );
FOR(shred_tile_cnt) fd_topob_tile_out( topo, "repair", 0UL, "repair_shred", i );
/**/ fd_topob_tile_out( topo, "sign", 0UL, "sign_ping", 0UL );

FOR(sign_tile_cnt-1) fd_topob_tile_out( topo, "repair", 0UL, "repair_sign", i );
FOR(sign_tile_cnt-1) fd_topob_tile_in ( topo, "sign", i+1, "metric_in", "repair_sign", i, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
FOR(sign_tile_cnt-1) fd_topob_tile_out( topo, "sign", i+1, "sign_repair", i );
FOR(sign_tile_cnt-1) fd_topob_tile_in ( topo, "repair", 0UL, "metric_in", "sign_repair", i, FD_TOPOB_UNRELIABLE, FD_TOPOB_POLLED );
/**/ fd_topob_tile_in ( topo, "repair", 0UL, "metric_in", "sign_ping", 0UL, FD_TOPOB_UNRELIABLE, FD_TOPOB_UNPOLLED );

/**/ fd_topob_tile_in ( topo, "gossip", 0UL, "metric_in", "sign_gossip", 0UL, FD_TOPOB_UNRELIABLE, FD_TOPOB_UNPOLLED );
/**/ fd_topob_tile_in ( topo, "gossip", 0UL, "metric_in", "sign_gossip", 0UL, FD_TOPOB_UNRELIABLE, FD_TOPOB_UNPOLLED );

if( 1 ) {
fd_topob_wksp( topo, "scap" );
Expand Down Expand Up @@ -451,7 +442,7 @@ static fd_location_info_t * location_table;
static fd_pubkey_t peers_copy[ FD_ACTIVE_KEY_MAX ];

static ulong
sort_peers_by_latency( fd_policy_peer_t * active_table, fd_peer_dlist_t * peers_dlist, fd_peer_t * peers_arr ) {
sort_peers_by_latency( fd_policy_peer_t * active_table, fd_peer_dlist_t * peers_dlist, fd_peer_dlist_t * peers_wlist, fd_peer_t * peers_arr ) {
ulong i = 0;
fd_peer_dlist_iter_t iter = fd_peer_dlist_iter_fwd_init( peers_dlist, peers_arr );
while( !fd_peer_dlist_iter_done( iter, peers_dlist, peers_arr ) ) {
Expand All @@ -461,6 +452,16 @@ sort_peers_by_latency( fd_policy_peer_t * active_table, fd_peer_dlist_t * peers_
if( FD_UNLIKELY( i >= FD_ACTIVE_KEY_MAX ) ) break;
iter = fd_peer_dlist_iter_fwd_next( iter, peers_dlist, peers_arr );
}
FD_LOG_NOTICE(( "best peers cnt: %lu", i ));
iter = fd_peer_dlist_iter_fwd_init( peers_wlist, peers_arr );
while( !fd_peer_dlist_iter_done( iter, peers_wlist, peers_arr ) ) {
fd_peer_t * peer = fd_peer_dlist_iter_ele( iter, peers_wlist, peers_arr );
if( FD_UNLIKELY( !peer ) ) break;
peers_copy[ i++ ] = peer->identity;
if( FD_UNLIKELY( i >= FD_ACTIVE_KEY_MAX ) ) break;
iter = fd_peer_dlist_iter_fwd_next( iter, peers_wlist, peers_arr );
}
FD_LOG_NOTICE(( "worst peers cnt: %lu", i ));

ulong peer_cnt = i;
for( uint i = 0; i < peer_cnt - 1; i++ ) {
Expand Down Expand Up @@ -494,12 +495,14 @@ print_peer_location_latency( fd_wksp_t * repair_tile_wksp, ctx_t * tile_ctx ) {
fd_policy_t * policy = fd_wksp_laddr ( repair_tile_wksp, policy_gaddr );
ulong peermap_gaddr = fd_wksp_gaddr_fast( tile_ctx->wksp, policy->peers.map );
ulong peerarr_gaddr = fd_wksp_gaddr_fast( tile_ctx->wksp, policy->peers.pool );
ulong peerlst_gaddr = fd_wksp_gaddr_fast( tile_ctx->wksp, policy->peers.dlist );
ulong peerlst_gaddr = fd_wksp_gaddr_fast( tile_ctx->wksp, policy->peers.best );
ulong peerwst_gaddr = fd_wksp_gaddr_fast( tile_ctx->wksp, policy->peers.worst );
fd_policy_peer_t * peers_map = (fd_policy_peer_t *)fd_wksp_laddr( repair_tile_wksp, peermap_gaddr );
fd_peer_dlist_t * peers_dlist = (fd_peer_dlist_t *)fd_wksp_laddr( repair_tile_wksp, peerlst_gaddr );
fd_peer_dlist_t * peers_wlist = (fd_peer_dlist_t *)fd_wksp_laddr( repair_tile_wksp, peerwst_gaddr );
fd_peer_t * peers_arr = (fd_peer_t *)fd_wksp_laddr( repair_tile_wksp, peerarr_gaddr );

ulong peer_cnt = sort_peers_by_latency( peers_map, peers_dlist, peers_arr );
ulong peer_cnt = sort_peers_by_latency( peers_map, peers_dlist, peers_wlist, peers_arr );
printf("\nPeer Location/Latency Information\n");
printf( "| %-46s | %-7s | %-8s | %-8s | %-7s | %12s | %s\n", "Pubkey", "Req Cnt", "Req B/s", "Rx B/s", "Rx Rate", "Avg Latency", "Location Info" );
for( uint i = 0; i < peer_cnt; i++ ) {
Expand Down Expand Up @@ -752,7 +755,7 @@ repair_cmd_fn_profiler_mode( args_t * args,
printf(" Repair Peers: %lu\n", repair_metrics[ MIDX( COUNTER, REPAIR, REQUEST_PEERS ) ] );
ulong slots_behind = turbine_slot0 > repair_metrics[ MIDX( COUNTER, REPAIR, REPAIRED_SLOTS ) ] ? turbine_slot0 - repair_metrics[ MIDX( COUNTER, REPAIR, REPAIRED_SLOTS ) ] : 0;
printf(" Repaired slots: %lu/%lu (slots behind: %lu)\n", repair_metrics[ MIDX( COUNTER, REPAIR, REPAIRED_SLOTS ) ], turbine_slot0, slots_behind );
if( !slots_behind ) { catchup_finished = 1; }
if( turbine_slot0 && !slots_behind ) { catchup_finished = 1; }
/* Print histogram buckets similar to Prometheus format */
print_histogram_buckets( repair_metrics,
MIDX( HISTOGRAM, REPAIR, SLOT_COMPLETE_TIME ),
Expand Down
11 changes: 2 additions & 9 deletions src/app/firedancer/topology.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -362,10 +362,8 @@ fd_topo_initialize( config_t * config ) {
/**/ fd_topob_link( topo, "gossip_sign", "gossip_sign", 128UL, 2048UL, 1UL ); /* TODO: Where does 2048 come from? Depth probably doesn't need to be 128 */
/**/ fd_topob_link( topo, "sign_gossip", "sign_gossip", 128UL, sizeof(fd_ed25519_sig_t), 1UL ); /* TODO: Depth probably doesn't need to be 128 */

FOR(sign_tile_cnt-1) fd_topob_link( topo, "repair_sign", "repair_sign", 128UL, FD_REPAIR_MAX_PREIMAGE_SZ, 1UL );
FOR(sign_tile_cnt-1) fd_topob_link( topo, "sign_repair", "sign_repair", 128UL, sizeof(fd_ed25519_sig_t), 1UL );
/**/ fd_topob_link( topo, "ping_sign", "repair_sign", 128UL, FD_REPAIR_MAX_PREIMAGE_SZ, 1UL );
/**/ fd_topob_link( topo, "sign_ping", "sign_repair", 128UL, sizeof(fd_ed25519_sig_t), 1UL );
FOR(sign_tile_cnt-1) fd_topob_link( topo, "repair_sign", "repair_sign", 128UL, FD_REPAIR_MAX_PREIMAGE_SZ, 1UL );
FOR(sign_tile_cnt-1) fd_topob_link( topo, "sign_repair", "sign_repair", 128UL, sizeof(fd_ed25519_sig_t), 1UL );

/**/ fd_topob_link( topo, "send_sign", "send_sign", 128UL, FD_TXN_MTU, 1UL ); /* TODO: Depth probably doesn't need to be 128 */
/**/ fd_topob_link( topo, "sign_send", "sign_send", 128UL, sizeof(fd_ed25519_sig_t), 1UL ); /* TODO: Depth probably doesn't need to be 128 */
Expand Down Expand Up @@ -580,15 +578,10 @@ fd_topo_initialize( config_t * config ) {
/**/ fd_topob_tile_out( topo, "sign", 0UL, "sign_shred", i );
}

/**/ fd_topob_tile_in ( topo, "sign", 0UL, "metric_in", "ping_sign", 0UL, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
/**/ fd_topob_tile_out( topo, "repair", 0UL, "ping_sign", 0UL );
/**/ fd_topob_tile_out( topo, "sign", 0UL, "sign_ping", 0UL );

FOR(sign_tile_cnt-1UL) fd_topob_tile_out( topo, "repair", 0UL, "repair_sign", i );
FOR(sign_tile_cnt-1UL) fd_topob_tile_in ( topo, "sign", i+1UL, "metric_in", "repair_sign", i, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
FOR(sign_tile_cnt-1UL) fd_topob_tile_out( topo, "sign", i+1UL, "sign_repair", i );
FOR(sign_tile_cnt-1UL) fd_topob_tile_in ( topo, "repair", 0UL, "metric_in", "sign_repair", i, FD_TOPOB_UNRELIABLE, FD_TOPOB_POLLED ); /* This link is polled because the signing requests are asynchronous */
/**/ fd_topob_tile_in ( topo, "repair", 0UL, "metric_in", "sign_ping", 0UL, FD_TOPOB_UNRELIABLE, FD_TOPOB_UNPOLLED );

/**/ fd_topob_tile_in ( topo, "sign", 0UL, "metric_in", "send_sign", 0UL, FD_TOPOB_RELIABLE, FD_TOPOB_POLLED );
/**/ fd_topob_tile_out( topo, "send", 0UL, "send_sign", 0UL );
Expand Down
2 changes: 1 addition & 1 deletion src/discof/forest/fd_forest.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,7 +79,7 @@ struct __attribute__((aligned(128UL))) fd_forest_blk {

fd_forest_blk_idxs_t code[fd_forest_blk_idxs_word_cnt]; /* code shred idxs */
long first_shred_ts; /* timestamp of first shred rcved in slot != complete_idx */
long first_req_ts; /* timestamp of first request received in slot != complete_idx */
long first_req_ts; /* tick of first request received in slot != complete_idx */
uint turbine_cnt; /* number of shreds received from turbine */
uint repair_cnt; /* number of data shreds received from repair */
uint recovered_cnt; /* number of shreds recovered from reedsol recovery */
Expand Down
54 changes: 38 additions & 16 deletions src/discof/repair/fd_policy.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,15 +30,17 @@ fd_policy_new( void * shmem, ulong dedup_max, ulong peer_max, ulong seed ) {
void * dedup_lru = FD_SCRATCH_ALLOC_APPEND( l, fd_policy_dedup_lru_align(), fd_policy_dedup_lru_footprint ( ) );
void * peers = FD_SCRATCH_ALLOC_APPEND( l, fd_policy_peer_map_align(), fd_policy_peer_map_footprint ( lg_peer_max ) );
void * peers_pool = FD_SCRATCH_ALLOC_APPEND( l, fd_peer_pool_align(), fd_peer_pool_footprint ( peer_max ) );
void * peers_dlist = FD_SCRATCH_ALLOC_APPEND( l, fd_peer_dlist_align(), fd_peer_dlist_footprint ( ) );
void * peers_best = FD_SCRATCH_ALLOC_APPEND( l, fd_peer_dlist_align(), fd_peer_dlist_footprint ( ) );
void * peers_worst = FD_SCRATCH_ALLOC_APPEND( l, fd_peer_dlist_align(), fd_peer_dlist_footprint ( ) );
FD_TEST( FD_SCRATCH_ALLOC_FINI( l, fd_policy_align() ) == (ulong)shmem + footprint );

policy->dedup.map = fd_policy_dedup_map_new ( dedup_map, dedup_max, seed );
policy->dedup.pool = fd_policy_dedup_pool_new( dedup_pool, dedup_max );
policy->dedup.lru = fd_policy_dedup_lru_new ( dedup_lru );
policy->peers.map = fd_policy_peer_map_new ( peers, lg_peer_max );
policy->peers.pool = fd_peer_pool_new ( peers_pool, peer_max );
policy->peers.dlist = fd_peer_dlist_new ( peers_dlist );
policy->peers.best = fd_peer_dlist_new ( peers_best );
policy->peers.worst = fd_peer_dlist_new ( peers_worst );
policy->iterf.ele_idx = ULONG_MAX;
policy->turbine_slot0 = ULONG_MAX;
policy->tsreset = 0;
Expand Down Expand Up @@ -72,8 +74,11 @@ fd_policy_join( void * shpolicy ) {
policy->dedup.lru = fd_policy_dedup_lru_join ( policy->dedup.lru );
policy->peers.map = fd_policy_peer_map_join ( policy->peers.map );
policy->peers.pool = fd_peer_pool_join ( policy->peers.pool );
policy->peers.dlist = fd_peer_dlist_join ( policy->peers.dlist );
policy->peers.iter = fd_peer_dlist_iter_fwd_init( policy->peers.dlist, policy->peers.pool );
policy->peers.best = fd_peer_dlist_join ( policy->peers.best );
policy->peers.worst = fd_peer_dlist_join ( policy->peers.worst );

policy->peers.select.iter = fd_peer_dlist_iter_fwd_init( policy->peers.worst, policy->peers.pool );
policy->peers.select.stage = 0;

return policy;
}
Expand Down Expand Up @@ -127,7 +132,7 @@ dedup_next( fd_policy_t * policy, ulong key, long now ) {
fd_policy_dedup_map_ele_insert ( dedup->map, ele, dedup->pool );
fd_policy_dedup_lru_ele_push_tail( dedup->lru, ele, dedup->pool );
}
if( FD_LIKELY( now < ele->req_ts + (long)80e6 ) ) {
if( FD_LIKELY( now < ele->req_ts + (long)FD_POLICY_DEDUP_TIMEOUT ) ) {
return 1;
}
ele->req_ts = now;
Expand Down Expand Up @@ -156,13 +161,19 @@ passes_throttle_threshold( fd_policy_t * policy, fd_forest_blk_t * ele ) {

fd_pubkey_t const *
fd_policy_peer_select( fd_policy_t * policy ) {
fd_peer_dlist_t * dlist = policy->peers.dlist;
fd_peer_t * pool = policy->peers.pool;
if( FD_UNLIKELY( fd_peer_dlist_iter_done( policy->peers.iter, dlist, pool ) ) ) {
policy->peers.iter = fd_peer_dlist_iter_fwd_init( dlist, pool );
fd_peer_dlist_t * best_dlist = policy->peers.best;
fd_peer_dlist_t * worst_dlist = policy->peers.worst;
fd_peer_t * pool = policy->peers.pool;

fd_peer_dlist_t * dlist = bucket_stages[policy->peers.select.stage] == FD_POLICY_LATENCY_BEST ? best_dlist : worst_dlist;

while( FD_UNLIKELY( fd_peer_dlist_iter_done( policy->peers.select.iter, dlist, pool ) ) ) {
policy->peers.select.stage = (policy->peers.select.stage + 1) % (sizeof(bucket_stages) / sizeof(uint));
dlist = bucket_stages[policy->peers.select.stage] == FD_POLICY_LATENCY_BEST ? best_dlist : worst_dlist;
policy->peers.select.iter = fd_peer_dlist_iter_fwd_init( dlist, pool );
}
fd_peer_t * select = fd_peer_dlist_iter_ele( policy->peers.iter, dlist, pool );
policy->peers.iter = fd_peer_dlist_iter_fwd_next( policy->peers.iter, dlist, pool );
fd_peer_t * select = fd_peer_dlist_iter_ele( policy->peers.select.iter, dlist, pool );
policy->peers.select.iter = fd_peer_dlist_iter_fwd_next( policy->peers.select.iter, dlist, pool );
return &select->identity;
}

Expand Down Expand Up @@ -294,7 +305,7 @@ fd_policy_peer_insert( fd_policy_t * policy, fd_pubkey_t const * key, fd_ip4_por
fd_peer_t * peer_ele = fd_peer_pool_ele_acquire( policy->peers.pool );
peer->pool_idx = fd_peer_pool_idx( policy->peers.pool, peer_ele );
peer_ele->identity = *key;
fd_peer_dlist_ele_push_tail( policy->peers.dlist, peer_ele, policy->peers.pool );
fd_peer_dlist_ele_push_tail( policy->peers.worst, peer_ele, policy->peers.pool );
return peer;
}
return NULL;
Expand All @@ -312,11 +323,14 @@ fd_policy_peer_remove( fd_policy_t * policy, fd_pubkey_t const * key ) {
fd_peer_t * peer_ele = fd_peer_pool_ele( policy->peers.pool, peer->pool_idx );
fd_policy_peer_map_remove( policy->peers.map, peer );

if( FD_UNLIKELY( policy->peers.iter == fd_peer_pool_idx( policy->peers.pool, peer_ele ) ) ) {
if( FD_UNLIKELY( policy->peers.select.iter == fd_peer_pool_idx( policy->peers.pool, peer_ele ) ) ) {
/* In general removal during iteration is safe, except when the iterator is on the peer to be removed. */
policy->peers.iter = fd_peer_dlist_iter_fwd_next( policy->peers.iter, policy->peers.dlist, policy->peers.pool );
fd_peer_dlist_t * dlist = policy->peers.select.stage == FD_POLICY_LATENCY_BEST ? policy->peers.best : policy->peers.worst;
policy->peers.select.iter = fd_peer_dlist_iter_fwd_next( policy->peers.select.iter, dlist, policy->peers.pool );
}
fd_peer_dlist_ele_remove( policy->peers.dlist, peer_ele, policy->peers.pool );

fd_peer_dlist_t * bucket = fd_policy_peer_latency_bucket( policy, peer->total_lat, peer->res_cnt );
fd_peer_dlist_ele_remove( bucket, peer_ele, policy->peers.pool );
fd_peer_pool_ele_release( policy->peers.pool, peer_ele );
return 1;
}
Expand All @@ -332,14 +346,22 @@ fd_policy_peer_request_update( fd_policy_t * policy, fd_pubkey_t const * to ) {
}

void
fd_policy_peer_response_update( fd_policy_t * policy, fd_pubkey_t const * to, long rtt ) {
fd_policy_peer_response_update( fd_policy_t * policy, fd_pubkey_t const * to, long rtt /* ns */ ) {
fd_policy_peer_t * peer = fd_policy_peer_query( policy, to );
if( FD_LIKELY( peer ) ) {
long now = fd_tickcount();
fd_peer_dlist_t * prev_bucket = fd_policy_peer_latency_bucket( policy, peer->total_lat, peer->res_cnt );
peer->res_cnt++;
if( FD_UNLIKELY( peer->first_resp_ts == 0 ) ) peer->first_resp_ts = now;
peer->last_resp_ts = now;
peer->total_lat += rtt;
fd_peer_dlist_t * new_bucket = fd_policy_peer_latency_bucket( policy, peer->total_lat, peer->res_cnt );

if( prev_bucket != new_bucket ) {
fd_peer_t * peer_ele = fd_peer_pool_ele( policy->peers.pool, peer->pool_idx );
fd_peer_dlist_ele_remove ( prev_bucket, peer_ele, policy->peers.pool );
fd_peer_dlist_ele_push_tail( new_bucket, peer_ele, policy->peers.pool );
}
}
}

Expand Down
Loading