diff --git a/test/nvmf/shutdown/shutdown.sh b/test/nvmf/shutdown/shutdown.sh
index d8938f25a..f247a1871 100755
--- a/test/nvmf/shutdown/shutdown.sh
+++ b/test/nvmf/shutdown/shutdown.sh
@@ -23,6 +23,30 @@ if [ -z $NVMF_FIRST_TARGET_IP ]; then
 	exit 0
 fi
 
+function waitforio() {
+	# $1 = RPC socket
+	if [ -z "$1" ]; then
+		exit 1
+	fi
+	# $2 = bdev name
+	if [ -z "$2" ]; then
+		exit 1
+	fi
+	local ret=1
+	local i
+	for (( i = 10; i != 0; i-- )); do
+		read_io_count=$($rpc_py -s $1 get_bdevs_iostat -b $2 | jq -r '.[1].num_read_ops')
+		# A few I/O will happen during initial examine.  So wait until at least 100 I/O
+		#  have completed to know that bdevperf is really generating the I/O.
+		if [ $read_io_count -ge 100 ]; then
+			ret=0
+			break
+		fi
+		sleep 0.25
+	done
+	return $ret
+}
+
 timing_enter shutdown
 timing_enter start_nvmf_tgt
 # Start up the NVMf target in another process
@@ -91,8 +115,7 @@ perfpid=$!
 waitforlisten $perfpid /var/tmp/bdevperf.sock
 $rpc_py -s /var/tmp/bdevperf.sock wait_subsystem_init
 
-# Sleep for a few seconds to allow I/O to begin
-sleep 5
+waitforio /var/tmp/bdevperf.sock Nvme1n1
 
 # Kill bdevperf half way through
 killprocess $perfpid
@@ -114,8 +137,7 @@ $rpc_py -s /var/tmp/bdevperf.sock wait_subsystem_init
 # Expand the trap to clean up bdevperf if something goes wrong
 trap "process_shm --id $NVMF_APP_SHM_ID; killprocess $pid; kill -9 $perfpid; nvmfcleanup; nvmftestfini $1; exit 1" SIGINT SIGTERM EXIT
 
-# Sleep for a few seconds to allow I/O to begin
-sleep 5
+waitforio /var/tmp/bdevperf.sock Nvme1n1
 
 # Kill the target half way through
 killprocess $pid