summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--docs/developer/design/multicast.rst278
-rw-r--r--docs/developer/spec/multicast.rst190
2 files changed, 468 insertions, 0 deletions
diff --git a/docs/developer/design/multicast.rst b/docs/developer/design/multicast.rst
new file mode 100644
index 00000000..89422fe6
--- /dev/null
+++ b/docs/developer/design/multicast.rst
@@ -0,0 +1,278 @@
+Detailed Design
+===============
+
+Protocol Design
+---------------
+
+1. All Protocol headers are 1 byte long or align to 4 bytes.
+2. Packet size should not exceed above 1500(MTU) bytes including UDP/IP header and should
+be align to 4 bytes. In future, MTU can be modified larger than 1500(Jumbo Frame) through
+cmd line option to enlarge the data throughput.
+
+/* Packet header definition (align to 4 bytes) */
+struct packet_ctl {
+ uint32_t seq; // packet seq number start from 0, unique in server life cycle.
+ uint32_t crc; // checksum
+ uint32_t data_size; // payload length
+ uint8_t data[0];
+};
+
+/* Buffer info definition (align to 4 bytes) */
+struct buffer_ctl {
+ uint32_t buffer_id; // buffer seq number start from 0, unique in server life cycle.
+ uint32_t buffer_size; // payload total length of a buffer
+ uint32_t packet_id_base; // seq number of the first packet in this buffer.
+ uint32_t pkt_count; // number of packet in this buffer, 0 means EOF.
+};
+
+
+3. 1-byte-long header definition
+
+Signals such as the four below are 1 byte long, to simplify the receive process(since it
+cannot be spitted ).
+
+#define CLIENT_READY 0x1
+#define CLIENT_REQ 0x2
+#define CLIENT_DONE 0x4
+#define SERVER_SENT 0x8
+
+Note: Please see the collaboration diagram for their meanings.
+
+4. Retransmission Request Header
+
+/* Retransmition Request Header (align to 4 bytes) */
+struct request_ctl {
+ uint32_t req_count; // How many seqs below.
+ uint32_t seqs[0]; // packet seqs.
+};
+
+5. Buffer operations
+
+void buffer_init(); // Init the buffer_ctl structure and all(say 1024) packet_ctl
+structures. Allocate buffer memory.
+long buffer_fill(int fd); // fill a buffer from fd, such as stdin
+long buffer_flush(int fd); // flush a buffer to fd, say stdout
+struct packet_ctl *packet_put(struct packet_ctl *new_pkt);// put a packet to a buffer
+and return a free memory slot for the next packet.
+struct packet_ctl *packet_get(uint32_t seq);// get a packet data in buffer by
+indicating the packet seq.
+
+
+How to sync between server threads
+----------------------------------
+
+If children's aaa() operation need to wait the parents's init() to be done, then do it
+literally like this:
+
+ UDP Server
+ TCP Server1 = spawn( )----> TCP Server1
+ init()
+ TCP Server2 = spawn( )-----> TCP Server2
+ V(sem)----------------------> P(sem) // No child any more
+ V(sem)---------------------> P(sem)
+ aaa() // No need to V(sem), for no child
+ aaa()
+
+If parent's send() operation need to wait the children's ready() done, then do it
+literally too, but is a reverse way:
+
+ UDP Server TCP Server1 TCP Server2
+ // No child any more
+ ready() ready()
+ P(sem) <--------------------- V(sem)
+ P(sem) <------------------ V(sem)
+ send()
+
+Note that the aaa() and ready() operations above run in parallel. If this is not the
+case due to race condition, the sequence above can be modified into this below:
+
+ UDP Server TCP Server1 TCP Server2
+ // No child any more
+ ready()
+ P(sem) <--------------------- V(sem)
+ ready()
+ P(sem) <------------------- V(sem)
+ send()
+
+
+In order to implement such chained/zipper sync pattern, a pair of semaphores is
+needed between the parent and the child. One is used by child to wait parent , the
+other is used by parent to wait child. semaphore pair can be allocated by parent
+and pass the pointer to the child over spawn() operation such as pthread_create().
+
+/* semaphore pair definition */
+struct semaphores {
+ sem_t wait_parent;
+ sem_t wait_child;
+};
+
+Then the semaphore pair can be recorded by threads by using the semlink struct below:
+struct semlink {
+ struct semaphores *this; /* used by parent to point to the struct semaphores
+ which it created during spawn child. */
+ struct semaphores *parent; /* used by child to point to the struct
+ semaphores which it created by parent */
+};
+
+chained/zipper sync API:
+
+void sl_wait_child(struct semlink *sl);
+void sl_release_child(struct semlink *sl);
+void sl_wait_parent(struct semlink *sl);
+void sl_release_parent(struct semlink *sl);
+
+API usage is like this.
+
+Thread1(root parent) Thread2(child) Thread3(grandchild)
+sl_wait_parent(noop op)
+sl_release_child
+ +---------->sl_wait_parent
+ sl_release_child
+ +-----------> sl_wait_parent
+ sl_release_child(noop op)
+ ...
+ sl_wait_child(noop op)
+ + sl_release_parent
+ sl_wait_child <-------------
+ + sl_release_parent
+sl_wait_child <------------
+sl_release_parent(noop op)
+
+API implementation:
+
+void sl_wait_child(struct semlink *sl)
+{
+ if (sl->this) {
+ P(sl->this->wait_child);
+ }
+}
+
+void sl_release_child(struct semlink *sl)
+{
+ if (sl->this) {
+ V(sl->this->wait_parent);
+ }
+}
+
+void sl_wait_parent(struct semlink *sl)
+{
+ if (sl->parent) {
+ P(sl->parent->wait_parent);
+ }
+}
+
+void sl_release_parent(struct semlink *sl)
+{
+ if (sl->parent) {
+ V(sl->parent->wait_child);
+ }
+}
+
+Client flow chart
+-----------------
+See Collaboration Diagram
+
+UDP thread flow chart
+---------------------
+See Collaboration Diagram
+
+TCP thread flow chart
+---------------------
+
+
+S_INIT --- (UDP initialized) ---> S_ACCEPT --- (accept clients) --+
+ |
+ /----------------------------------------------------------------/
+ V
+S_PREP --- (UDP prepared abuffer)
+ ^ |
+ | \--> S_SYNC --- (clients ClIENT_READY)
+ | |
+ | \--> S_SEND --- (clients CLIENT_DONE)
+ | |
+ | V
+ \---------------(bufferctl.pkt_count != 0)-----------------------+
+ |
+ V
+ exit() <--- (bufferctl.pkt_count == 0)
+
+
+TCP using poll and message queue
+--------------------------------
+
+TCP uses poll() to sync with client's events as well as output event from itself, so
+that we can use non-block socket operations to reduce the latency. POLLIN means there
+are message from client and POLLOUT means we are ready to send message/retransmission
+packets to client.
+
+poll main loop pseudo code:
+void check_clients(struct server_status_data *sdata)
+{
+ poll_events = poll(&(sdata->ds[1]), sdata->ccount - 1, timeout);
+
+ /* check all connected clients */
+ for (sdata->cindex = 1; sdata->cindex < sdata->ccount; sdata->cindex++) {
+ ds = &(sdata->ds[sdata->cindex]);
+ if (!ds->revents) {
+ continue;
+ }
+
+ if (ds->revents & (POLLERR|POLLHUP|POLLNVAL)) {
+ handle_error_event(sdata);
+ } else if (ds->revents & (POLLIN|POLLPRI)) {
+ handle_pullin_event(sdata); // may set POLLOUT into ds->events
+ // to trigger handle_pullout_event().
+ } else if (ds->revents & POLLOUT) {
+ handle_pullout_event(sdata);
+ }
+ }
+}
+
+For TCP, since the message from client may not complete and send data may be also
+interrupted due to non-block fashion, there should be one send message queue and a
+receive message queue on the server side for each client (client do not use non-block
+operations).
+
+TCP message queue definition:
+
+struct tcpq {
+ struct qmsg *head, *tail;
+ long count; /* message count in a queue */
+ long size; /* Total data size of a queue */
+};
+
+TCP message queue item definition:
+
+struct qmsg {
+ struct qmsg *next;
+ void *data;
+ long size;
+};
+
+TCP message queue API:
+
+// Allocate and init a queue.
+struct tcpq * tcpq_queue_init(void);
+
+// Free a queue.
+void tcpq_queue_free(struct tcpq *q);
+
+// Return queue length.
+long tcpq_queue_dsize(struct tcpq *q);
+
+// queue new message to tail.
+void tcpq_queue_tail(struct tcpq *q, void *data, long size);
+
+// queue message that cannot be sent currently back to queue head.
+void tcpq_queue_head(struct tcpq *q, void *data, long size);
+
+// get one piece from queue head.
+void * tcpq_dequeue_head(struct tcpq *q, long *size);
+
+// Serialize all pieces of a queue, and move it out of queue, to ease the further
+//operation on it.
+void * tcpq_dqueue_flat(struct tcpq *q, long *size);
+
+// Serialize all pieces of a queue, do not move it out of queue, to ease the further
+//operation on it.
+void * tcpq_queue_flat_peek(struct tcpq *q, long *size);
diff --git a/docs/developer/spec/multicast.rst b/docs/developer/spec/multicast.rst
new file mode 100644
index 00000000..ba314d3a
--- /dev/null
+++ b/docs/developer/spec/multicast.rst
@@ -0,0 +1,190 @@
+Requirement
+===========
+1. When deploying a large OPNFV/OpenStack cluster, we would like to take the advantage of UDP
+multicast to prevent the network bottleneck when distributing Kolla container from one
+Installer Server to all target hosts by using unicast.
+
+2. When it comes to auto scaling (extension) of compute nodes, use unicast is acceptable, since
+the number of nodes in this condition is usually small.
+
+The basic step to introduce multicast to deployment is:
+a. Still setup the monopolistic docker registry server on Daisy server as a failsafe.
+b. Daisy server, as the multicast server, prepares the image file to be transmitted, and count
+how many target hosts(as the multicast clients)that should receive the image file
+simultaneously.
+c. Multicast clients tell the multicast server about ready to receive the image.
+d. Multicast server transmits image over UDP multicast channel.
+e. Multicast clients report success after received the whole image.
+f. Setup docker registry server on each target hosts based upon received docker image.
+g. Setup Kolla ansible to use 127.0.0.1 as the registry server IP so that the real docker
+container retrieving network activities only take place inside target hosts.
+
+
+Design
+======
+
+Methods to achieve
+------------------
+
+TIPC
+++++
+
+TIPC or its wrapper such as ZeroMQ is good at multicast, but it is not suitable as an
+installer:
+1. The default TIPC kernel module equipped by CentOS7(kernel verison 3.10) is NOT stable
+especially in L3 multicast(although we can use L2 multicast, but the network will be limited to
+L2). If errors happen, it is hard for us to recover a node from kernel panic.
+
+2. TIPC's design is based on a stable node cluster environment, esp in Lossless Ethernet. But
+the real environment is generally not in that case. When multicast is broken, Installer should
+switch to unicast, but TIPC currently do not have such capability.
+
+Top level design
+----------------
+1. There are two kinds of thread on the server side, one is UDP multicast thread the other is
+TCP sync/retransmit thread. There will be more than one TCP threads since one TCP thread can
+only serve a limited client (say 64~128) in order to limit the CPU load and unicast retransmit
+network usage.
+
+2. There is only one thread on client side.
+
+3. All the packets that a client lost during UDP multicast will be request by client to the TCP
+thread and resend by using TCP unicast, if unicast still cannot deliver the packets successfully,
+the client will failback to using the monopolistic docker registry server on Daisy server as a
+failsafe option.
+
+4. Each packet needs checksum.
+
+
+UDP Server Design (runs on Daisy Server)
+----------------------------------------
+
+1. Multicast group IP and Port should be configurable, as well as the interface that will be
+used as the egress of the multicast packets. The user will pass the interface's IP as the
+handle to find the egress.
+
+2. Image data to be sent is passed to server through stdin.
+
+3. Consider the size of image is large (xGB), the server cannot pre-allocate whole buffer to
+hold all image at once. Besides, since the data is from stdin and the actual length is
+unpredictable. So the server should split the data into small size buffers and send to the
+clients one by one. Furthermore, buffer shall be divided into packets which size is MTU
+including the UDP/IP header. Then the buffer size can be , for example 1024 * MTU including the
+UDP/IP header.
+
+4. After sending one buffer to client the server should stop and get feedback from client to
+see if all clients have got all packets in that buffer. If any clients lost any buffer, client
+should request the server to resend packets from a more stable way(TCP).
+
+5. when got the EOF from stdin, server should send a buffer which size is 0 as an EOF signal to
+the client to let it know about the end of sending.
+
+
+TCP Server Design (runs on Daisy Server)
+----------------------------------------
+
+1. All TCP server threads and the only one UDP thread share one process. The UDP thread is the
+parent thread, and the first TCP thread is the child, while the second TCP thread is the
+grandchild, and so on. Thus, for each TCP thread, there is only one parent and at most one
+child.
+
+2. TCP thread accepts the connect request from client. The number of client is predefined by
+server cmdline parameter. Each TCP thread connect with at most ,say 64 clients, if there are
+more clients to be connected to, then a child TCP thread is spawned by the parent.
+
+3. Before UDP thread sending any buffer to client, all TCP threads should send UDP multicast
+IP/Port information to their clients beforehand.
+
+4. During each buffer sending cycle, TCP threads send a special protocol message to tell
+clients about the size/id of the buffer and id of each packet in it. After getting
+acknowledgements from all clients, TCP threads then signal the UDP thread to start
+multicasting buffer over UDP. After multicasting finished, TCP threads notifies clients
+multicast is done, and wait acknowledgements from clients again. If clients requests
+retransmission, then it is the responsibility of TCP threads to resend packets over unicast.
+If no retransmission needed, then clients should signal TCP threads that they are ready for
+the next buffer to come.
+
+5. Repeat step 4 if buffer size is not 0 in the last round, otherwise, TCP server shutdown
+connection and exit.
+
+
+Server cmdline usage example
+----------------------------
+
+./server <local_ip> <number_of_clients> [port] < kolla_image.tgz
+
+<local_ip> is used here to specify the multicast egress interface. But which interface will be
+used by TCP is leaved to route table to decide.
+<number_of_clients> indicates the number of clients , thus the number of target hosts which
+need to receive the image.
+[port] is the port that will be used by both UDP and TCP. Default value can be used if user
+does not provide it.
+
+
+Client Design(Target Host side)
+--------------------------------
+
+1. Each target hosts has only one client process.
+
+2. Client connect to TCP server according to the cmdline parameters right after start up.
+
+3. After connecting to TCP server, client first read from TCP server the multicast group
+information which can be used to create the multicast receive socket then.
+
+4. During each buffer receiving cycle, the client first read from TCP server the buffer info,
+prepare the receive buffer, and acknowledge the TCP server that it is ready to receive. Then,
+client receive buffer from the multicast socket until TCP server notifying the end of
+multicast. By compare the buffer info and the received packets, the client knows whether to
+send the retransmission request or not and whether to wait retransmission packet or not.
+After all packets are received from UDP/TCP, the client eventually flush buffer to stdout
+and tells the TCP server about ready to receive the next buffer.
+
+5. Repeat step 4 if buffer size is not 0 in the last round, otherwise, client shutdowns
+connection and exit.
+
+Client cmdline usage example
+----------------------------
+
+./client <local_ip> <server_ip> [port] > kolla_image.tgz
+
+<local_ip> is used here to specify the multicast ingress interface. But which interface
+will be used by TCP is leaved to route table to decide.
+<server_ip> indicates the TCP server IP to be connected to.
+[port] is the port that will be used by both connect to TCP server and receive multicast
+data.
+
+
+Collaboration diagram among UDP Server, TCP Server(illustrate only one TCP thread)
+and Clients:
+
+
+UDP Server TCP Server Client
+ | | |
+init mcast group
+init mcast send socket
+ ---------------------------------->
+ accept clients
+ <------------------------connet------------------
+ --------------------send mcast group info------->
+ <----------------------------------
+ state = PREP
+do {
+read data from stdin
+prepare one buffer
+ ----------------------------------->
+ state = SYNC
+ -------------------send buffer info-------------->
+ <----------------------send ClIENT_READY-----------
+ <----------------------------------
+ state = SEND
+
+ ================================================send buffer over UDP multicast======>
+ ----------------------------------->
+ -----------------------send SERVER_SENT----------->
+ [<-------------------send CLIENT_REQUEST----------]
+ [--------------send buffer over TCP unicast------>]
+ flush buffer to stdout
+ <-------------------send CLIENT_DONE---------------
+ <----------------------------------
+ state = PREP
+while (buffer.len != 0)