Tutorial 3: IPCP statistics
For this tutorial, you should have a local layer, a normal layer and a ping server registered in the normal layer. You will need to have the FUSE libraries installed and Ouroboros compiled with FUSE support. We will show you how to get some statistics from the network layer which is exported by the IPCPs at /tmp/ouroboros (this mountpoint can be set at compile time):
$ tree /tmp/ouroboros
/tmp/ouroboros/
|-- ipcpd-normal.13569
| |-- dt
| | |-- 0
| | |-- 1
| | `-- 65
| `-- lsdb
| |-- 416743497.465922905
| |-- 465922905.416743497
| |-- dt.465922905
| `-- mgmt.465922905
`-- ipcpd-normal.4363
|-- dt
| |-- 0
| |-- 1
| `-- 65
`-- lsdb
|-- 416743497.465922905
|-- 465922905.416743497
|-- dt.416743497
`-- mgmt.416743497
6 directories, 14 files
There are two filesystems, one for each normal IPCP. Currently, it shows information for two components: data transfer and the link-state database. The data transfer component lists flows on known flow descriptors. The flow allocator component will usually be on fd 0 and the directory (DHT). There is a single (N-1) data transfer flow on fd 65 that the IPCPs can use to send data (these fd’s will usually not be the same). The routing component sees that data transfer flow as two unidirectional links. It has a management flow and data transfer flow to its neighbor. Let’s have a look at the data transfer flow in the network:
$ cat /tmp/ouroboros/ipcpd-normal.13569/dt/65
Flow established at: 2018-03-07 18:47:43
Endpoint address: 465922905
Queued packets (rx): 0
Queued packets (tx): 0
Qos cube 0:
sent (packets): 4
sent (bytes): 268
rcvd (packets): 3
rcvd (bytes): 298
local sent (packets): 4
local sent (bytes): 268
local rcvd (packets): 3
local rcvd (bytes): 298
dropped ttl (packets): 0
dropped ttl (bytes): 0
failed writes (packets): 0
failed writes (bytes): 0
failed nhop (packets): 0
failed nhop (bytes): 0
<no traffic on other qos cubes>
The above output shows the statistics for the data transfer component of the IPCP that enrolled into the layer. It shows the time the flow was established, the endpoint address and the number of packets that are in the incoming and outgoing queues. Then it lists packet statistics per QoS cube. It sent 4 packets, and received 3 packets. All the packets came from local sources (internal components of the IPCP) and were delivered to local destinations. Let’s have a look where they went.
$ cat /tmp/ouroboros/ipcpd-normal.13569/dt/1
Flow established at: 2018-03-07 18:47:43
Endpoint address: flow-allocator
Queued packets (rx): 0
Queued packets (tx): 0
<no packets on this flow>
There is no traffic on fd 0, which is the flow allocator component. This will only be used when higher layer applications will use this normal layer. Let’s have a look at fd 1.
$ cat /tmp/ouroboros/ipcpd-normal.13569/dt/1
Flow established at: 2018-03-07 18:47:43
Endpoint address: dht
Queued packets (rx): 0
Queued packets (tx): 0
Qos cube 0:
sent (packets): 3
sent (bytes): 298
rcvd (packets): 0
rcvd (bytes): 0
local sent (packets): 0
local sent (bytes): 0
local rcvd (packets): 6
local rcvd (bytes): 312
dropped ttl (packets): 0
dropped ttl (bytes): 0
failed writes (packets): 0
failed writes (bytes): 0
failed nhop (packets): 2
failed nhop (bytes): 44
<no traffic on other qos cubes>
The traffic for the directory (DHT) is on fd1. Take note that this is from the perspective of the data transfer component. The data transfer component sent 3 packets to the DHT, these are the 3 packets it received from the data transfer flow. The data transfer component received 6 packets from the DHT. It only sent 4 on fd 65. 2 packets failed because of nhop. This is because the forwarding table was being updated from the routing table. Let’s send some traffic to the oping server.
$ oping -n oping_server -i 0
Pinging oping_server with 64 bytes of data:
64 bytes from oping_server: seq=0 time=0.547 ms
...
64 bytes from oping_server: seq=999 time=0.184 ms
--- oping_server ping statistics ---
1000 SDUs transmitted, 1000 received, 0% packet loss, time: 106.538 ms
rtt min/avg/max/mdev = 0.151/0.299/2.269/0.230 ms
This sent 1000 packets to the server. Let’s have a look at the flow allocator:
$ cat /tmp/ouroboros/ipcpd-normal.13569/dt/0
Flow established at: 2018-03-07 18:47:43
Endpoint address: flow-allocator
Queued packets (rx): 0
Queued packets (tx): 0
Qos cube 0:
sent (packets): 1
sent (bytes): 59
rcvd (packets): 0
rcvd (bytes): 0
local sent (packets): 0
local sent (bytes): 0
local rcvd (packets): 1
local rcvd (bytes): 51
dropped ttl (packets): 0
dropped ttl (bytes): 0
failed writes (packets): 0
failed writes (bytes): 0
failed nhop (packets): 0
failed nhop (bytes): 0
<no traffic on other qos cubes>
The flow allocator has sent and received a message: a request and a response for the flow allocation between the oping client and server. The data transfer flow will also have additional traffic:
$ cat /tmp/ouroboros/ipcpd-normal.13569/dt/65
Flow established at: 2018-03-07 18:47:43
Endpoint address: 465922905
Queued packets (rx): 0
Queued packets (tx): 0
Qos cube 0:
sent (packets): 1013
sent (bytes): 85171
rcvd (packets): 1014
rcvd (bytes): 85373
local sent (packets): 13
local sent (bytes): 1171
local rcvd (packets): 14
local rcvd (bytes): 1373
dropped ttl (packets): 0
dropped ttl (bytes): 0
failed writes (packets): 0
failed writes (bytes): 0
failed nhop (packets): 0
failed nhop (bytes): 0
This shows the traffic from the oping application. The additional traffic (the oping sent 1000, the flow allocator 1 and the DHT previously sent 3) is additional DHT traffic (the DHT periodically updates). Also note that the traffic reported on the link includes the FRCT and data transfer headers which in the default configuration is 20 bytes per packet.
This concludes tutorial 3.