1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
|
---
title: "Flow allocation"
author: "Dimitri Staessens"
#description: protocols
date: 2020-01-17
weight: 30
draft: false
description: >
The most important concept in Ouroboros
---
Arguably the most important concept to grasp in Ouroboros is flow
allocation.[^1] It is the process by which a pair of programs agree to
start sending and receiving data, and the interface to the network. A
flow is always unicast, thus between a source program and a
destination program, and is always established from the source. Flows
are provided by unicast layers, and the endpoints of the flows are
accessible for reading and writing by the requesting processes using
an identifier called a _flow descriptor_. Think of a file descriptor
but just for Ouroboros flows. Maybe one important thing to keep in
mind: in Ouroboros terminology, a flow does not imply ordering or
reliable transfer. It just denotes the network resources inside a
layer that are needed for forwarding packets from a source to a
destination in a best effort way.
{{<figure width="60%" src="/docs/concepts/fa_1.jpg">}}
The figure above gives an example. There are 2 systems, and each
system has an Ouroboros IRMd and a unicast IPCP. These IPCPs work
together to create a logical "layer". System 1 runs a "client"
program, System 2 runs a "server" program.
We are going to explain in some detail the steps that Ourobros takes
to establish a flow between the "client" and "server" program so they
can communicate.
The three subcomponents inside the IPCP that are of interest to us are
the Directory (DIR), the Flow Allocator (FA) and the Data Transfer
component (DT).
The DT component is at the heart of the network functionality in the
layer. It is a protocol machine responsible for forwarding packets and
maintains a forwarding table that maps destination addresses to lower
layer flows. [^2] The name of the DT is what is generally considered the
"address" of the IPCP. [^3] In the example, IPCP 1 has address 720,
and IPCP 2 address 1000. If DT 720 receives a packet for DT 1000, it
will know how to forward it to 1000 and vice versa. I will not go into
the details of how routing information is distributed, suffice to say
it's similar in operation to the IS-IS protocol. The only other thing
that is of current interest is the protocol format of the DT
component. The DT protocol has [5 fields](../protocols) [^4]:
```
DST | TTL | QoS | ECN | EID |
```
To understand the flow allocation procedure, we need to consider only
2 of these fields, the destination address (DST), and the endpoint ID
(EID). I will denote the relevant packet header information for DT in
the format __DST:EID__ So, __1000:78__ would indicate a packet
destined for IPCP 2 with EID 78. EID's are a bit like a tcp port, but
they are not well-known (i.e. there is no IANA in Ouroboros). The flow
allocation process will assign the EIDs.
The directory (DIR) component keeps a mapping of registered hashes to
DT names (addresses). For the server application to be reachable over
a layer, the DIR component in its IPCPs will have to know this
mapping. In our example, the server, which is named server is known by
the layer to be at location 1000. The interface to register a name is
actually using hashes, so "server" is hashed (by default an SHA3-256
hash) to _d19778d2_[^5] and a mapping (_d19778d2_, 1000) is kept in the
directory. The default implementation for the DIR component in the
Ouroboros IPCP is a Distributed Hash Table (DHT) based on the Kademlia
protocol.
The third subcomponent in the IPCP that is relevant here -- the most
important one -- is the Flow Allocator (FA). This component is
responsible for implementing and managing requested flows. It is also
responsible for congestion control.
For establishing a flow, it needs to establish some shared state
between the two endpoints. A (bidirectional) flow is fully identified
in a layer by a 4-tuple (A1,X,A2,Y) containing two addresses and two
EIDs, in our example A1=720 and A2=1000). This 4-tuple needs to be
known at both endpoints to identify where to send the packets it
receives from the higher-layer application (the client), and to
deliver packets that it reads from a lower layer flow. The flow
allocation protocol is responsible to send this information. It is a
request-response protocol. The flow allocator is identified by the DT
component as EID 0. So, all packets in the layer with DT header
__DST:0__ are delivered to the flow allocator inside the destintation
IPCP.
When the source FA in IPCP 1 receives a request for a flow to
"server", it will query its DIR for _d197782_ and receive 1000 as the
response and it will generate an EID (X) for the flow. Let's assume
X=75. The flow allocation request protocol message from FA 1 to FA 2
looks like __1000:0:REQ:720:75:d19778d2__, and when FA 2 received this
message, it will generate its EID, let's say 81 and send the following
response to FA 1: __720:0:RESP:75:81__. REQ and RESP are internal
codes to identify a request and reponse (0 and 1 respectively). From
this small exchange both flow allocators can now identify the flow.
Finally, there is the IRMd in each system. The IRMd should be seen as
part of the operating system. One of its tasks is to map process IDs
(PIDs) of a process to names. In our example above, the IRMd in System
two will have a mapping that maps _d19778d2_ to the PID of
"server". When the "server" program calls the Ouroboros
_flow\_accept()_ routine, the IRMd knows that when there is an
incoming flow allocation request, the "server" process can handle
it. Populating this mapping in the IRMd is a process we call _binding_
a name to a process.
Let's now go step-by-step through the full flow allocation process in
the example above.
{{<figure width="60%" src="/docs/concepts/fa_2.jpg">}}
The first few steps are shown in the figure above. The client
application requests a flow to "server" to the Ouroboros IRMd using
the _flow\_alloc()_ call __(1)__. Now the IRMd will ask the layers in
the system if they know that name "server", indirectly by using the
SHA3-256 hash, _d1977d2_ __(2)__. The hash algorithm that a layer uses
is configurable, and the IRMd is informed of the hash algorithm to use
when an IPCP joins a layer (at bootstrap or enrollment). In our case,
the layer shown will respond to the query with "True" __(3)__, (multiple
layers can respond true, and then the IRMd will choose one, usually
the "lowest" in rank). Note that the results of these queries can be
cached locally in the IRMd to speed up the process.
So, now that the IRMd knows that the layer in the figure knows the
destination program, it can send a flow allocation request to the
layer. But first, it will start creating some local resources: the
flow endpoint, indicated by a flow_id (FID) __(4)__. It contains a set of
ring-buffers in shared memory that contain pointer information on
where to read/write the next packet. The FID will be in _PENDING_
state [^6].
{{<figure width="60%" src="/docs/concepts/fa_3.jpg">}}
When the FID resources are ready, the IRMd sends a _FLOW\_ALLOC_ to
the IPCP with the pending FID as endpoint __(5)__. The FA in IPCP 1
will create a _flow descriptor_ for this flow [^7], let's say 75
__(6)__. All packets that are written by the IPCP to fd 75 can be read
from FID 9. Now, a couple of paragraphs ago I mentioned that the FA
will generate an EID for the flow. In the implementation, the EID for
the flow equals the fd. So packets coming from within the layer with
EID 75 will be written to this flow.
This is the point where the FA will do the flow allocation protocol
exchange already described above. The destination hash is resolved
from the directory to the destination IPCP address, 1000, and the
following flow allocation request message is sent over the DT
component to the destination IPCP: __1000:0:REQ:720:75:d19778d2__ __(7)__.
{{<figure width="60%" src="/docs/concepts/fa_4.jpg">}}
We can now turn our attention to System 2, which receives this request
message on IPCP 2. The DT header contains __1000:0__ which has the
correct address (1000) and EID 0, which indicates the packet should be
delivered to the flow allocator. So the FA interprets the following
information from the received packet: There is a flow allocation
request for the hash _d19778d2_ coming from source 720 on remote EID
75.
Now it send a _FLOW\_ALLOC\_REQ()_ message to the IRMd __(8)__. The
IRMd has in its process table an entry that says that there is a
process that listens to this hash. It will create a flow endpoint, for
instance with FID=16 __(9)__ and respond to the IPCPd that the flow is
accepted with FID=16 __(10)__. The _flow\_accept()_ call on the server
side will return with an fd=71 that points to the FID 9. From this
point on, the server can use the flow __(11)__.
The flow allocator in IPCP 2 can now complete its enpoint
configuration. It will create a mapping [S_EID -> R_ADDR, R_EID], in
this case [81 -> 720, 75]. So all packets that it reads from EID 81
will get a header __720:75__ from the DT component __(12)__. It will
now complete the flow allocation protocol and send a response message
that flow allocation succeeded. The contents of this message is
__720:0:RES:75:81__ __(13)__. This concludes all operations on the
server side.
{{<figure width="60%" src="/docs/concepts/fa_5.jpg">}}
Back to the client. The FA in System 1 receives the packet, and from
EID 0 knows it is for the flow allocator, which gets its last piece of
information: the remote EID for the flow, 81. It can now create its
own mapping, [75->1000, 81] __(14)__ and respond to its IRMd that the
flow is created __(15)__. The IRMd will change the state of the flow
from _PENDING_ to _ALLOCATED_ __(16)__ and the _flow\_accept()_ call
on the client program will return with an FD for the flow. The flow is
now allocated.
So, from now on, communication between the server and the client is
pretty straightforward. Data is written to some shared memory in an
buffer that allows for some space to prepend headers and append
CRCs. To avoid memory copies, pointers to these locations are passed
over the ringbuffers in the flow endpoints to the IPCP, which reads
the pointers, adds headers in the right location, and then uses the
same procedure to pass it onto the next layer towards the destination.
The translation of the header is an O(1) lookup on the send side, and
a nop on the receiver side (since FD == EID and it's passed in the
packet).
[^1]: This concept is also present in RINA, but there are differences. This text only applies to Ouroboros.
[^2]: This is a recursive network, adjancencies in layer N are implemented as flows in layer N - 1.
[^3]: If there is one DT, it is what is usually considered a "flat" address. More complex addressing schemes are accomplished by having more of these DT components inside one IPCP. But this would lead us too far. It is described in more detail in the paper.
[^4]: I will explain QoS in a different post.
[^5]: In full: d19778d2e34a1e3ddfc04b48c94152cced725d741756b131543616d20f250f31.
[^6]: Note that the _flow\_alloc()_ call __1__ is currently blocking. Asynchronous allocation implementation is on the TODO list.
[^7]: All this mapping of fd's is done by the library that is used by all Ouroboros programs.
|