Improve error monitoring

net/net.sh

@@ -13,7 +13,7 @@ usage() {
     echo "Error: $*"
   fi
   cat <<EOF
-usage: $0 [start|stop]
+usage: $0 [start|stop|restart|sanity] [command-specific options]
 
 Operate a configured testnet
 
@@ -114,16 +114,11 @@ build() {
 
 common_start_setup() {
   declare ipAddress=$1
-  declare logFile="$2"
-
-  (
-    set -x
-    test -d "$SOLANA_ROOT"
-    ssh "${sshOptions[@]}" "$ipAddress" "mkdir -p ~/solana ~/.cargo/bin"
-    rsync -vPr -e "ssh ${sshOptions[*]}" \
-      "$SOLANA_ROOT"/{fetch-perf-libs.sh,scripts,net,multinode-demo} \
-      "$ipAddress":~/solana/
-  ) >> "$logFile" 2>&1
+  test -d "$SOLANA_ROOT"
+  ssh "${sshOptions[@]}" "$ipAddress" "mkdir -p ~/solana ~/.cargo/bin"
+  rsync -vPrc -e "ssh ${sshOptions[*]}" \
+    "$SOLANA_ROOT"/{fetch-perf-libs.sh,scripts,net,multinode-demo} \
+    "$ipAddress":~/solana/
 }
 
 startLeader() {
@@ -131,18 +126,17 @@ startLeader() {
   declare logFile="$2"
   echo "--- Starting leader: $leaderIp"
 
-  common_start_setup "$ipAddress" "$logFile"
-
   # Deploy local binaries to leader.  Validators and clients later fetch the
   # binaries from the leader.
   (
     set -x
+    common_start_setup "$ipAddress" || exit 1
     case $deployMethod in
     snap)
-      rsync -vPr -e "ssh ${sshOptions[*]}" "$snapFilename" "$ipAddress:~/solana/solana.snap"
+      rsync -vPrc -e "ssh ${sshOptions[*]}" "$snapFilename" "$ipAddress:~/solana/solana.snap"
       ;;
     local)
-      rsync -vPr -e "ssh ${sshOptions[*]}" "$SOLANA_ROOT"/farf/bin/* "$ipAddress:~/.cargo/bin/"
+      rsync -vPrc -e "ssh ${sshOptions[*]}" "$SOLANA_ROOT"/farf/bin/* "$ipAddress:~/.cargo/bin/"
       ;;
     *)
       usage "Internal error: invalid deployMethod: $deployMethod"
@@ -151,7 +145,11 @@ startLeader() {
 
     ssh "${sshOptions[@]}" -n "$ipAddress" \
       "./solana/net/remote/remote-node.sh $deployMethod leader $leaderIp $expectedNodeCount \"$nodeSetupArgs\" \"$RUST_LOG\""
-  ) >> "$logFile" 2>&1
+  ) >> "$logFile" 2>&1 || {
+    cat "$logFile"
+    echo "^^^ +++"
+    exit 1
+  }
 }
 
 startValidator() {
@@ -160,8 +158,8 @@ startValidator() {
 
   echo "--- Starting validator: $leaderIp"
   (
-    common_start_setup "$ipAddress" /dev/stdout
     set -x
+    common_start_setup "$ipAddress"
     ssh "${sshOptions[@]}" -n "$ipAddress" \
       "./solana/net/remote/remote-node.sh $deployMethod validator $leaderIp $expectedNodeCount \"$nodeSetupArgs\" \"$RUST_LOG\""
   ) >> "$netLogDir/validator-$ipAddress.log" 2>&1 &
@@ -173,14 +171,18 @@ startValidator() {
 startClient() {
   declare ipAddress=$1
   declare logFile="$2"
-  echo "--- Starting client: $leaderIp"
-  common_start_setup "$ipAddress" "$logFile"
+  echo "--- Starting client: $ipAddress"
 
   (
     set -x
+    common_start_setup "$ipAddress"
     ssh "${sshOptions[@]}" -f "$ipAddress" \
       "./solana/net/remote/remote-client.sh $deployMethod $leaderIp $expectedNodeCount \"$RUST_LOG\""
-  ) >> "$logFile" 2>&1
+  ) >> "$logFile" 2>&1 || {
+    cat "$logFile"
+    echo "^^^ +++"
+    exit 1
+  }
 }
 
 sanity() {
@@ -191,7 +193,7 @@ sanity() {
     # shellcheck disable=SC2029 # remote-client.sh args are expanded on client side intentionally
     ssh "${sshOptions[@]}" "$leaderIp" \
       "./solana/net/remote/remote-sanity.sh $sanityExtraArgs"
-  )
+  ) || exit 1
 }
 
 start() {

net/remote/remote-client.sh

@@ -1,4 +1,4 @@
-#!/bin/bash -e
+#!/bin/bash -ex
 
 cd "$(dirname "$0")"/../..
 
@@ -22,12 +22,13 @@ scripts/install-earlyoom.sh
 
 case $deployMethod in
 snap)
-  rsync -vPr "$leaderIp:~/solana/solana.snap" .
+  rsync -vPrc "$leaderIp:~/solana/solana.snap" .
   sudo snap install solana.snap --devmode --dangerous
   rm solana.snap
 
   nodeConfig="\
     leader-ip=$leaderIp \
+    default-metrics-rate=1 \
     metrics-config=$SOLANA_METRICS_CONFIG \
     rust-log=$RUST_LOG \
   "
@@ -39,9 +40,10 @@ snap)
 local)
   PATH="$HOME"/.cargo/bin:"$PATH"
   export USE_INSTALL=1
+  export SOLANA_DEFAULT_METRICS_RATE=1
   export RUST_LOG
 
-  rsync -vPr "$leaderIp:~/.cargo/bin/solana*" ~/.cargo/bin/
+  rsync -vPrc "$leaderIp:~/.cargo/bin/solana*" ~/.cargo/bin/
   solana_bench_tps="multinode-demo/client.sh $leaderIp:~/solana"
   ;;
 *)
@@ -49,14 +51,16 @@ local)
   exit 1
 esac
 
-scripts/oom-monitor.sh  > oom-monitor.log 2>&1 &
+scripts/oom-monitor.sh > oom-monitor.log 2>&1 &
 
 while true; do
   echo "=== Client start: $(date)" >> client.log
-  clientCommand="$solana_bench_tps --num-nodes $numNodes --loop -s 600 --sustained -t threadCount"
+  clientCommand="$solana_bench_tps --num-nodes $numNodes --seconds 600 --sustained --threads $threadCount"
   echo "$ $clientCommand" >> client.log
 
+  set +e
   $clientCommand >> client.log 2>&1
+  set -e
 
   $metricsWriteDatapoint "testnet-deploy,name=$netBasename clientexit=1"
   echo Error: bench-tps should never exit | tee -a client.log

net/remote/remote-node.sh

@@ -28,11 +28,12 @@ scripts/install-earlyoom.sh
 case $deployMethod in
 snap)
   SECONDS=0
-  rsync -vPr "$leaderIp:~/solana/solana.snap" .
+  rsync -vPrc "$leaderIp:~/solana/solana.snap" .
   sudo snap install solana.snap --devmode --dangerous
 
   commonNodeConfig="\
     leader-ip=$leaderIp \
+    default-metrics-rate=1 \
     metrics-config=$SOLANA_METRICS_CONFIG \
     rust-log=$RUST_LOG \
     setup-args=$setupArgs \
@@ -65,6 +66,7 @@ local)
   PATH="$HOME"/.cargo/bin:"$PATH"
   export USE_INSTALL=1
   export RUST_LOG
+  export SOLANA_DEFAULT_METRICS_RATE=1
   if [[ -e /dev/nvidia0 ]]; then
     export SOLANA_CUDA=1
   fi
@@ -80,7 +82,7 @@ local)
     ./multinode-demo/leader.sh > leader.log 2>&1 &
     ;;
   validator)
-    rsync -vPr "$leaderIp:~/.cargo/bin/solana*" ~/.cargo/bin/
+    rsync -vPrc "$leaderIp:~/.cargo/bin/solana*" ~/.cargo/bin/
 
     # shellcheck disable=SC2086 # Don't want to double quote "$setupArgs"
     ./multinode-demo/setup.sh -t validator -p $setupArgs

net/remote/remote-sanity.sh

@@ -100,10 +100,13 @@ fi
 echo "--- $leaderIp: validator sanity"
 if $validatorSanity; then
   (
+    set -ex -o pipefail
     ./multinode-demo/setup.sh -t validator
-    set -e pipefail
-    timeout 10s ./multinode-demo/validator.sh "$leaderIp" 2>&1 | tee validator.log
-  )
+    timeout 10s ./multinode-demo/validator.sh "$leaderIp" "$leaderIp:8001" 2>&1 | tee validator.log
+  ) || {
+    exitcode=$?
+    [[ $exitcode -eq 124 ]] || exit $exitcode
+  }
   wc -l validator.log
   if grep -C100 panic validator.log; then
     echo "^^^ +++"