pgmdump
-------
View all packets like tcpdump, but updated to full spec and allow dump of payload.
pgmtop
------
Dump realtime packet statistics in a ncurses display, mix of top/htop/netop.
basic_send
----------
Send an ODATA packet and terminate.
Accept string payload and network parameters on command line.
Send to multicast or send to unicast AFI.
IPv4/6.
Define optional session start, late join tags.
spm_idle
--------
Idle in an event loop sending out SPM packets.
stream_send
-----------
Send a constant stream of ODATA and SPM packets.
basic_http
----------
Simple embedded web server
basic_recv
----------
Listen to packets indicating data loss, view details through web interface.
basic_container
---------------
Test performance of glib containers for fast allocating for a dynamic transmit window.
basic_txw
---------
Test performance of a basic transmit window implementation.
nak_txw
-------
Test performance of random access to packets inside the window.
stream_send_with_nak
--------------------
Respond to NAK's with RDATA.
basic_recv_with_nak
-------------------
Listen to packets and send NAK's to rebuild data.
dumpif
------
Display all IP based interfaces and basic details.
testif
------
Test various combinations of network specification.
sw_calc
-------
Basic calculation tests of wrap-around sliding windows and a leading edge.
basic_recv_with_rxw
-------------------
Listen to packets buffered with a receive window.
test_cpu_timers
---------------
Calculate drift between processors, cores, and hyper-threads.
pgmsend
--------
basic_send updated to use transmit window.
pgmrecv
--------
basic_recv_with_rxw without web interface, primary displays messages from pgmsend.
syncrecv
--------
pgmrecv implemented outside GLib with a synchronous coding paradigm.
pgmping
-------
Dual mode: one to send fixed messages like pgmsend and listen for response, two to listen for
messages and reply.
block_send
----------
Send APDUs over ODATA.
(pgmrecv can receive APDUs)
test_rs
-------
Test 8-bit symbol Reed Solomon encoding and decoding with errors & erasures.
test_fec
--------
Test fec creation and recovery.
send_with_fec
--------------------
Send APDUs over ODATA with FEC.
Scenarios to reproduce
**********************
- Packet loss in stream causing NAK generation.
- Link saturation in sending causing API feedback.
- Link peak stable speed.
- Maxium NAK generation to determine NCF/RDATA throughput.
- Corrupt packets with invalid IP checksum (? generate IP HDR in sender)
- Corrupt packets with invalid PGM checksum.
- Invalid packet values.
- NAK to NCF latency.
- NAK to RDATA latency.
- Publish bandwidth: total, per packet type, payload, per recipient (?)
- Subscribe bandwidth: total, per packet type, payload, per publisher (?)
- Restarting a session with similar or dissimilar sequence numbering.
Outstanding questions
*********************
- Is it faster to use chunks containing multiple packets instead of one MTU
per packet. Does aligning with system page size matter?
- Can g_timers be dropped easily for gettimeofday and avoid floating point math? Possible
to pass timer upstream with contiguous data for easy access.
- Can time evaluation be dropped to at most once per main loop event?
- Does code work 32 bit and is it optimal?
- Should trash stacks be monitored and controlled externally? For example, clearing
up after bursts or administrative control.
- Should trash stacks have a upper limit to then free to the slice allocator?
- Should lost packets be managed as ranges or individual sequence numbers, how
does each method affect performance?
* The initial draft of PGM included OPT_RANGE option for NAKs to specify a range of lost
packets, this was replaced in the final draft with NAK lists. Some research hints that
ranges are suitable:
http://www.isoc.org/inet2001/CD_proceedings/T54/T54.htm
http://tools.ietf.org/html/draft-speakman-pgm-spec-01
- Are place holders necessary? Can state timeouts be managed without a per sequence number
object? For example by the next data object, or an extra object for an ncf extended window:
note that nak packet generation should easily dwarfs time spent unless advantage is taken
of the additional 62 naks in a opt_nak_list. Caution has to be taken with the cost of
splitting a range when a packet is inserted in the middle, although idealy it should
be sequential from the trailing edge.
- Is it better to have send sockets per transport, or shared, bound to each interface?
- Cost of sharing state helper lists between receive windows? On culling a peer the lists
have to be purged. Saves iterating the hash list of receivers.
- Encapsulated UDP lists two ports, 3305 for broadcast, 3306 for unicast, how is this
supposed to map to regular PGM, and why the split?
Basic TODO list
***************
- Ensure standardised error handling and status reporting.
- Implement mechanism for data-loss feedback.
- OPT_SYN & OPT_FIN on sending side.
- Shared trash stacks between multiple receive windows (contexts).
- Shared trash stacks between transmit and receive windows.
- FEC: compatibility with SmartPGM FEC, MS FEC?
- NAK backoff learning.
- Full conformance testing. (nak backoffs remaining)
- Unit testing.
- System testing with valgrind.
- Performance testing with oprofile.
- Basic DLR.
- Implement PGM sender (only) support thread?
- eventfd instead of pipes for recent Linux kernels.
Optionals
*********
- Some form of broadcast statistics for passive monitoring.
- (fix) BCH Reed-Solomon codec as possibly faster alternative, albeit incompatible with Microsoft.
- Recommendations as per the RMT working group of the IETF for AL-FEC codecs: Raptor codes, LDPC-
staircase and LDPC-triangle codes.
- XDR based messaging format as example of binary encoded messaging.
