diff --git a/python-ecosys/iperf3/iperf3.py b/python-ecosys/iperf3/iperf3.py
index 05d69f774..14be439b5 100644
--- a/python-ecosys/iperf3/iperf3.py
+++ b/python-ecosys/iperf3/iperf3.py
@@ -71,7 +71,8 @@ def fmt_size(val, div):
 
 class Stats:
     def __init__(self, param):
-        self.pacing_timer_us = param["pacing_timer"] * 1000
+        # iperf3 clients before 3.2 do not send the pacing timer.
+        self.pacing_timer_us = param.get("pacing_timer", 1000) * 1000
         self.udp = param.get("udp", False)
         self.reverse = param.get("reverse", False)
         self.running = False
@@ -217,8 +218,10 @@ def _transfer(udp, reverse, addr, s_data, buf, udp_last_send, udp_packet_id, udp
                 stats.add_bytes(n)
     else:
         if reverse:
-            recvninto(s_data, buf)
-            n = len(buf)
+            # The socket is non-blocking, so this reads only what is available.  The
+            # sender is not required to end the stream on a block boundary, so waiting
+            # here for a full block could block forever and miss TEST_END.
+            n = recvinto(s_data, buf) or 0
         else:
             n = s_data.send(buf)
         stats.add_bytes(n)
@@ -233,7 +236,9 @@ def server_once():
     s_listen = socket.socket(ai[0], socket.SOCK_STREAM)
     s_listen.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
     s_listen.bind(ai[-1])
-    s_listen.listen(1)
+    # The client opens its parallel streams all at once, so leave room to queue them.
+    # lwIP has 5 TCP pcbs by default, enough for the control connection and 4 streams.
+    s_listen.listen(4)
     s_ctrl, addr = s_listen.accept()
 
     # Read client's cookie
@@ -251,15 +256,22 @@ def server_once():
     if DEBUG:
         print(param)
     reverse = param.get("reverse", False)
+    parallel = param.get("parallel", 1)
 
     # Ask to create streams
     s_ctrl.sendall(bytes([CREATE_STREAMS]))
 
     if param.get("tcp", False):
-        # Accept stream
-        s_data, addr = s_listen.accept()
-        print("Accepted connection:", addr)
-        recvn(s_data, COOKIE_SIZE)
+        # Accept streams, the client opens one connection per parallel stream
+        s_data = []
+        for _ in range(parallel):
+            s, addr = s_listen.accept()
+            print("Accepted connection:", addr)
+            recvn(s, COOKIE_SIZE)
+            if not reverse:
+                # Receive without blocking, see _transfer().
+                s.setblocking(False)
+            s_data.append(s)
         udp = False
         udp_packet_id = 0
         udp_interval = None
@@ -267,10 +279,11 @@ def server_once():
     elif param.get("udp", False):
         # Close TCP connection and open UDP "connection"
         s_listen.close()
-        s_data = socket.socket(ai[0], socket.SOCK_DGRAM)
-        s_data.bind(ai[-1])
-        data, addr = s_data.recvfrom(4)
-        s_data.sendto(struct.pack("<I", 987654321), addr)
+        s = socket.socket(ai[0], socket.SOCK_DGRAM)
+        s.bind(ai[-1])
+        data, addr = s.recvfrom(4)
+        s.sendto(struct.pack("<I", 987654321), addr)
+        s_data = [s]
         udp_interval = 1000000 * 8 * param["len"] // param["bandwidth"]
         udp = True
         udp_packet_id = 0
@@ -287,10 +300,9 @@ def server_once():
     # Read data, and wait for client to send TEST_END
     poll = select.poll()
     poll.register(s_ctrl, select.POLLIN)
-    if reverse:
-        poll.register(s_data, select.POLLOUT)
-    else:
-        poll.register(s_data, select.POLLIN)
+    for s in s_data:
+        poll.register(s, select.POLLOUT if reverse else select.POLLIN)
+    stream_bytes = [0] * len(s_data)
     stats = Stats(param)
     stats.start()
     running = True
@@ -303,32 +315,34 @@ def server_once():
                     print(cmd_string.get(cmd, "UNKNOWN_COMMAND"))
                 if cmd == TEST_END:
                     running = False
-            elif pollable_is_sock(pollable, s_data):
-                udp_last_send, udp_packet_id = _transfer(
-                    udp,
-                    not reverse,
-                    addr,
-                    s_data,
-                    data_buf,
-                    udp_last_send,
-                    udp_packet_id,
-                    udp_interval,
-                    stats,
-                )
+                continue
+            # Find the stream that is ready, checking the first one first.
+            i = 0
+            while not pollable_is_sock(pollable, s_data[i]):
+                i += 1
+            nb = stats.nb0
+            udp_last_send, udp_packet_id = _transfer(
+                udp,
+                not reverse,
+                addr,
+                s_data[i],
+                data_buf,
+                udp_last_send,
+                udp_packet_id,
+                udp_interval,
+                stats,
+            )
+            stream_bytes[i] += stats.nb0 - nb
         stats.update()
 
-    # Need to continue writing so other side doesn't get blocked waiting for data
-    if reverse and not udp:
-        while True:
-            for pollable in poll.poll(0):
-                if pollable_is_sock(pollable, s_data):
-                    s_data.send(data_buf)
-                    break
-            else:
-                break
-
     stats.stop()
 
+    # Close the data streams now, as the reference implementation does.  A client that
+    # reads whole blocks would otherwise wait for the rest of one, and the unread data
+    # held for the streams can leave no room to receive the client's results.
+    for s in s_data:
+        s.close()
+
     # Ask to exchange results
     s_ctrl.sendall(bytes([EXCHANGE_RESULTS]))
 
@@ -348,15 +362,17 @@ def server_once():
         "congestion_used": "cubic",
         "streams": [
             {
-                "id": 1,
-                "bytes": stats.nb0,
+                # The reference implementation numbers its streams 1, 3, 4, 5, ...
+                "id": i + 2 if i else 1,
+                "bytes": stream_bytes[i],
                 "retransmits": 0,
                 "jitter": 0,
                 "errors": 0,
-                "packets": stats.np0,
+                "packets": 0 if i else stats.np0,
                 "start_time": 0,
                 "end_time": ticks_diff(stats.t3, stats.t0) * 1e-6,
             }
+            for i in range(len(s_data))
         ],
     }
     results = json.dumps(results)
@@ -371,7 +387,6 @@ def server_once():
     assert cmd == IPERF_DONE
 
     # Close all sockets
-    s_data.close()
     s_ctrl.close()
     s_listen.close()
 
@@ -488,6 +503,9 @@ def client(host, udp=False, reverse=False, bandwidth=10 * 1024 * 1024):
                         s_data = socket.socket(ai[0], socket.SOCK_STREAM)
                         s_data.connect(ai[-1])
                         s_data.sendall(cookie)
+                        if reverse:
+                            # Receive without blocking, see _transfer().
+                            s_data.setblocking(False)
                     buf = bytearray(urandom(param["len"]))
                 elif cmd == EXCHANGE_RESULTS:
                     # Close data socket now that server knows we are finished, to prevent it flooding us
diff --git a/python-ecosys/iperf3/manifest.py b/python-ecosys/iperf3/manifest.py
index d63c2d6d1..da1bf927b 100644
--- a/python-ecosys/iperf3/manifest.py
+++ b/python-ecosys/iperf3/manifest.py
@@ -1,3 +1,3 @@
-metadata(version="0.1.5", pypi="iperf3", pypi_publish="uiperf3")
+metadata(version="0.2.0", pypi="iperf3", pypi_publish="uiperf3")
 
 module("iperf3.py")