IOC connection time analysis

README.md

IOC connection time analysis

We will use caproto-shark to analyze CA network traffic. Note that the actual servers and clients involve may or may not be using caproto themselves; it does not matter.

1. Install caproto and pandas if they are not already installed.

   pip install caproto[standard] pandas
   

2. Start tcpdump to capture network traffic and write it to a file.

   sudo tcpdump -i <INTERFACE> -w some_network_traffic.pcap
   

   Use ifconfig to list the available interfaces. For example, if the servers in question are running on the localhost:

   sudo tcpdump -i lo -w some_network_traffic.pcap
   

3. Run some EPICS client code that will create CA connections with servers (e.g. run bsui).

4. Stop tcpdump with Ctrl+C.

5. Analyze the results using the script in this gist.

   python benchmark.py some_network_traffic.pcap
   

Example output:

0 channel creation requests were unanswered during this capture.


Raw data:
          PV  server_address             t        dt
0   simple:A   192.168.86.22  1.614185e+09  0.014851
1   simple:B   192.168.86.22  1.614185e+09  0.014851
2   simple:C   192.168.86.22  1.614185e+09  0.014851
3  rpi:color  192.168.86.245  1.614185e+09  0.150308


Aggregated stats by server address:
                count      mean  std       min       25%       50%       75%       max
server_address                                                                        
192.168.86.22     3.0  0.014851  0.0  0.014851  0.014851  0.014851  0.014851  0.014851
192.168.86.245    1.0  0.150308  NaN  0.150308  0.150308  0.150308  0.150308  0.150308

benchmark.py

"""
Analyze IOCs response times for channel creation (connection).

python benchmark.py some_network_traffic.pcap
"""
from collections import namedtuple
from caproto import CreateChanRequest, CreateChanResponse
from caproto.sync.shark import shark
import pandas


Pair = namedtuple("Pair", ["request", "response"])
Record = namedtuple("Record", ["PV", "server_address", "t", "dt"])


def match_request_and_response(parsed):
    "From a stream of parsed CA traffic, extract channel creation request/response pairs."
    unanswered_requests = {}
    pairs = []
    for item in parsed:
        command = item.command
        if isinstance(command, CreateChanRequest):
            unanswered_requests[command.cid] = item
        elif isinstance(command, CreateChanResponse):
            request_item = unanswered_requests.pop(command.cid, None)
            pairs.append(Pair(request_item, item))
    return pairs, list(unanswered_requests.values())


def build_record_from_request_and_response_pair(pair):
    "Extract PV name, server address, absolute time t, and request/response dt."
    request, response = pair
    record = Record(
        PV=request.command.name,
        server_address=request.dst,
        t=request.timestamp,
        dt=response.timestamp - request.timestamp
    )
    return record


def main(filepath):
    with open(filepath, "rb") as file:
        parsed = shark(file)
        pairs, unanswered = match_request_and_response(parsed)
    print(f"{len(unanswered)} channel creation requests were unanswered during this capture.")
    # TODO Tally the unanswered requests by server_address and print that.
    # The collections.Counter object would be useful there.
    records = [build_record_from_request_and_response_pair(pair) for pair in pairs]
    df = pandas.DataFrame.from_records(records, columns=Record._fields)
    server_stats = df.groupby("server_address")["dt"].describe()
    print("\n\nRaw data:")
    print(df)

    print("\n\nAggregated stats by server address:")
    print(server_stats)


if __name__ == "__main__":
    import sys

    main(sys.argv[1])