summaryrefslogtreecommitdiffstats
path: root/docs/developer/spec
diff options
context:
space:
mode:
authorZhijiang Hu <hu.zhijiang@zte.com.cn>2017-04-07 01:36:45 +0000
committerGerrit Code Review <gerrit@opnfv.org>2017-04-07 01:36:45 +0000
commitd24bd81e5689c5d8531581fb8e3604d6851cc094 (patch)
tree5701dd2eed93a794bd2457a144a367981930880d /docs/developer/spec
parent8f52576ce2063e1e742e1e6fafd66f6979bb2d10 (diff)
parenta051fc3bf0ea4cd589b663e974517633563a4ed8 (diff)
Merge "Add multicast spec"
Diffstat (limited to 'docs/developer/spec')
-rw-r--r--docs/developer/spec/multicast.rst190
1 files changed, 190 insertions, 0 deletions
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)