Skip to content
This repository
Newer
Older
100644 154 lines (123 sloc) 6.168 kb
0b8aa412 » BuzzTroll
2010-09-21 merged virga1 and master
1 ==========
2 Lantorrent
3 ==========
4
5 Lantorrent is a file distribution protocol integrated into the Nimbus
6 IaaS toolkit. It works as a means to multi-cast virtual machine images
7 to many backend nodes. The protocol is optimized for propagating
8 virtual machine images (typically large files) from a central repository
9 across a LAN to many virtual machine monitor nodes.
10
11 Lantorrent works best for the following scenarios:
12
13 1) large file transfers (VM images are typically measured in gigabytes)
14 2) local area switched network (typical for data center computer racks)
15 3) file recipients are willing peers.
16 4) many endpoints request the same file at roughly the same time
17
18 -----------------
19 Protocol overview
20 -----------------
21
22 When an endpoint wants a file it submits a request to a central agent.
23 This agent aggregates request for files so that they can be sent out in
24 an efficient single multi-cast session. Each request for a source file
25 is stored until either N request on that file have been made or N'
26 seconds have passed since the last request on that source file has been
27 made. This allows for a user to request a single file in several
28 unrelated session yet still have the file transfered in an efficient
29 multi-cast session.
30
31 Once N requests for a given source file have been made or N' seconds
32 have passed the destination set for the source file is determined. A
33 chain of destination endpoints is formed such that each node receives
34 from and sends to one other node. The first node receives from the
35 repository and send to a peer node, that peer node sends to another,
36 and so on until all receive the file. In this way all links of the
37 switch are utilized to send directly to another endpoint in the switch.
38 This results in the most efficient transfer on a LAN switched network.
39
40 Often times in a IaaS system a single network endpoint (VMM) will want
41 multiple copies of the same file. Each file is booted as a virtual
42 machine and that virtual machine will make distinct changes to that file
43 as it runs, thus it needs it own copy of the file. However that file
44 does not need to be transfered across the network more than once.
45 Lantorrent will send the file to each endpoint once and instruct that
46 endpoint to write it to multiple files if needed.
47
48 -----------------------------
49 Enabling Lantorrent in Nimbus
50 -----------------------------
51
52 Lan torrent is part of the nimbus distribution as of Nimbus 2.XXX.
53 However, due to system administrative overhead it is not enabled by
54 default. To enable Lantorrent in nimbus there are a few configurations
55 changes that must be made.
56
57 1) edit $NIMBUS_HOME/nimbus-setup.conf
58 change lantorrent.enabled: False -> lantorrent.enabled: True
59
60 2) edit $NIMBUS_HOME/services/etc/nimbus/workspace-service/other/common.conf
61 change the value of propagate.extraargs:
62 propagate.extraargs=$NIMBUS_HOME/lantorrent/bin/lt-request
63
64 be sure to expand $NIMBUS_HOME to its full and actual path.
65
66 3) xinetd
67 Lantorrent is run out of xinetd thus it must be installed on all VMMs.
68
69 4) install lantorrent on VMM
70 - recursively copy $NIMBUS_HOME/lantorrent to /opt/nimbus/lantorrent.
71 - run ./vmm-install.sh on each node
72 either run it as your workspace control user or specify the workspace
73 control user as the first and only argument to the script.
74
75 4.1) install lantorrent into xinetd
76 - the vmm-install,sh script creates the file lantorrent.inet. This
77 file is ready to be copied into /etc/xinetd.d/. Once this is done
78 restart xinetd (/etc/init.d/xinetd restart).
79
80 5) [optional] if the path to nimbus on the workspace control nodes (VMMs)
81 is not /opt/nimbus you will also need to edit a configuration file on
82 all backends.
83
84 In the file:
85 <workspace control path>/control/etc/workspace-control/propagation.conf
86
87 make sure the value of:
88
89 lantorrentexe: /opt/nimbus/bin/ltclient.sh
90
91 points to the proper location of you ltclinet.sh script. This should
92 be a simple matter of changing /opt/nimbus to the path where you chose
93 to install workspace control.
94
95 ----------------
96 Protocol details
97 ----------------
98
99 requests
100 --------
101
102 transfer
103 --------
104
105 A transfer from one peer to another is performed for forming a TCP
106 connection. A json header is first sent down the socket describing the
107 transfer. It describes the files where data must be written, and the
108 other endpoints where the data must be sent. As data comes in the
109 server will store it to one or more files and forward it on to another
110 endpoint. All of this information is contained in the header. The
111 header has the following format:
112
113 {
114 host
115 port
116 length
117 requests = [ {filename, id, rename}, ]
118 destinations =
119 [ {
120 host
121 port
122 requests = [ { filename, id, rename } ]
123 block_size
124 }, ]
125 }
126
127 When the server forms a connection with the next endpoint it will form
128 another header by removing the first entry of the destination list and
129 using the values contained in it as root level values in the header
130 described above. If the destination list is empty then the server is the
131 last in the chain and no forwarding is needed.
132
133 After the header is sent the binary payload (the contents of the file)
134 is sent down the socket.
135
136 The 'requests' value of the json header contains destination
137 information. Each filename in the list is a file where data should be
138 written. The value of id is the id of the request for that file. The
139 value of rename is a boolean which determines if the file should be
140 first written to a temporary location and then moved to the final
141 location once all of the data is received. This feature is helpful in
142 monitoring file completion.
143
144 security
145 --------
146
147 Each connection to a lantorrent peer comes with a json header. Directly
148 following that header is a line of text in the following format:
149 EOH : <signature>
150
151 The value of signature is the hmac signature of everything that came
152 before the literal EOH. The password for that signature is found in the
153 lt.ini file. Every endpoint and the master repo must have the same
154 value for that password.
155
156
Something went wrong with that request. Please try again.