List of TODO items (oldest first)

----------------------------------------------------------------------

Subject: Write enough unit tests for first release
Status: done

There are plenty of extended tests in the tests/ directory, but I need
to have some simpler unit test cases. Besides the usual use/use_ok
stuff, any check that's currently implemented as a die/croak/warn can
be converted into a unit test case.

Also, I need to write test cases that ensure that the "fast" xor
routine work in the same way as the "safe" one. I don't need too many
tests for this as there are more tests in the ctest/ directory.

----------------------------------------------------------------------

Subject: Gather a list of references
Status: done

Basically the wikipedia article and the main papers referenced from
there. Also link to Gwylim Ashley's python implementation.

----------------------------------------------------------------------

Subject: Investigate why the "asymptotically good in n" promise fails
Status: pending

The algorithm is supposed to be able to decode a message after
receiving a small number of check blocks over and above the number of
message/composite blocks. In my testing, this isn't working.

I may have to contact the papers' authors in order to find out what's
going wrong.

----------------------------------------------------------------------

Subject: Write "multi-speed" example program
Status: pending

One of the nice applications described in one of the papers I've read
is to have a sender transmitting over several different "channels",
each of which has a distinct bandwidth rate (no. of check blocks sent
per time unit). Listeners subscribe to one or more channels based on
their ability to receive check blocks sent over that channel. The idea
is that the sender doesn't have to receive explicit ACK/NAK messages
from nodes that have poor network connectivity: so long as any nodes
are subscribed to a lower-bandwidth channel, it simply continues
sending check blocks their way (and doesn't stop sending until it's
sent enough check blocks on that channel for the message to be
reconstructed).

As a variation of this, I want a program that will be testable over a
local network that is separated into two subnets. Assume that the
sender is on a high-speed subnet, so it can send at full speed to each
local machine on this subnet via multicast messaging. Also assume that
we have a router that can route multicast traffic to the second
subnet.  Assume that this second subnet is slower than the main subnet
(even if it's not, we can just pretend that it is). So we will send
check blocks on a separate multicast channel at a slower speed.

Besides joining a separate multicast channel for fast and slow
receivers, we also set the ttl ("time-to-live") parameter for each
multicast packet sent. We use a value of 1 for local (fast) network
peers, with a value of 2 for peers that live on the other side of the
router and have a slower network.

----------------------------------------------------------------------

Subject: get POD documentation into shape
Status: done

Some things may have changed since I started coding, so I have to
check what I've already written.

The Net::OnlineCode module needs to document what Online Codes are,
and include references. (done)

I also need to include a warning that the interfaces might change
between the initial release and subsequent releases. In particular,
I'm leaning towards changing how the Decoder class works. See the next
TODO for that.

----------------------------------------------------------------------

Subject: rewrite Decoder to be be more POE-friendly
Status: pending

When using POE as an event loop, we generally want each event handler
(callback) to take only a small time to complete. This way, we can do
a small bit of work and then return control back to POE's main event
loop, which can check for newly arrived network packets and so on.

As the code stands, whenever we receive a check block the resolve()
method tries to resolve as many auxiliary/message blocks as possible
before returning. Unfortunately, if this takes too long (because too
many new composite blocks have become decodable at once) then we may
not return to the POE main loop for quite a while, and we may
potentially miss receipt of several new check blocks as a result.

By limiting the amount of work done in the resolve to only decode at
most one composite block at a time (and arranging for any outstanding
decoding work to be scheduled for the next invocation), we will have
only a short delay each time we enter resolve(), and the calling
program can then interleave checks for newly-incoming check blocks
with the resolution step(s) without fear of dropping packets.


----------------------------------------------------------------------

Subject: Reduce memory requirements
Status: done

The algorithm is supposed to work best when there are a large number
(eg, in the order of 100,000 to 1 million) of check blocks being sent.
Unfortunately, as it stands, the program is a real memory hog when
many check blocks are being received. I need to look into reducing the
memory footprint, eg:

* free up useless object-local storage (eg, auxiliary mapping?)
* make sure to free memory where possible (eg, finished check blocks)
* possibly move to using more space-efficient structures for edges
* possibly move to C-based implementation for decoder

After trying (or considering) some of the above, I ended up changing
the algorithm so that it didn't automatically expand message/aux
blocks into the full list of check blocks they depend on. I also used
a profiler (nytprof) in parallel with this. The net result was an
order of magnitude increase in throughput and a similar decrease in
memory usage.

I haven't yet tested for 1m check blocks, but the tests (with 5,000
and 10,000 message blocks) seem promising.


----------------------------------------------------------------------

Subject: Make installing C/XS part of the code optional
Status: pending

Although I'm sure most people would prefer to have optimised C
routines when they can, not all platforms where Perl is installed will
have a C compiler. I'd like to make the module at least installable on
such platforms.

Related to this is making Net::OnlineCode::xor_strings call the
appropriate safe/fast routine by default. May be as simple as:

  our (*xor_strings) = safe_xor_strings|fast_xor_strings; #?

----------------------------------------------------------------------

Subject: get modules CPAN-ready
Status: done

The usual suspects:

* license details (including copyright year) (done)
* having useful unit tests (done)
* project-specific README (done)
* $VERSION in each file (done)
* sanity-check MANIFEST file (and try installing from dist tarfile) (done)
* POD included/up-to-date (done)
* test on a variety of platforms (x86/x86_64/ARM) (oops.. forgot)

Afterwards:

* check smoke tests online and fix any problems


----------------------------------------------------------------------

Subject: Fix smoke test errors
Status: done

Edited error outputs:

#   Failed test 'use Net::OnlineCode::Decoder;'
#   at t/05_Decoder.t line 4.
#     Tried to use 'Net::OnlineCode::Decoder'.
#     Error:  Type of arg 1 to push must be array (not hash element) at GraphDecoder.pm line 80, near "};"
# Type of arg 1 to push must be array (not array element) at GraphDecoder.pm line 235, near "$node;"
# Type of arg 1 to keys must be hash or array (not array element) at GraphDecoder.pm line 465, near "];"
# Type of arg 1 to push must be array (not hash element) at GraphDecoder.pm line 506, near "};"
# Type of arg 1 to push must be array (not hash element) at GraphDecoder.pm line 513, near "$nodelist;"
# Type of arg 1 to push must be array (not array element) at GraphDecoder.pm line 522, near "$node;"
# Type of arg 1 to push must be array (not hash element) at GraphDecoder.pm line 529, near "$new_hash;"
# Type of arg 1 to keys must be hash or array (not array element) at GraphDecoder.pm line 586, near "];"
# Type of arg 1 to keys must be hash or array (not array element) at GraphDecoder.pm line 637, near "];"
# Compilation failed in require at Decoder.pm line 9.
# BEGIN failed--compilation aborted at Decoder.pm line 9.
# Compilation failed in require at (eval 4) line 2.
# BEGIN failed--compilation aborted at (eval 4) line 2.
# Looks like you failed 1 test of 1.
t/05_Decoder.t ......... 

All of these are for Perl versions < 5.14. The problem is apparently
that I'm using hashrefs and arrayrefs and operating on them as if they
were hashes and arrays. This doesn't work in older versions of
perl. See perllol/Growing Your Own.

I'll install an old (5.004) version of Perl somewhere and see if I can
rework the code to use explicit dereferences. In the meantime, I've
marked all the places where the errors cropped up with the comment
"5.14".

Rather than test with older version of Perl (which proved impossible
with my distro without installing from source), I've just put the
appropriate @{} and %{} casts around the offending items.

The other problem that I can deduce from the cpantester reports is
that I was using 'integer-type int' in my Build.PL. This works fine
for "short int", "long int" and so on, but obviously fails for "int
int", which isn't a C type.

----------------------------------------------------------------------

Subject: Write a more memory-efficient Set module
Status: pending

None of the alternatives look like they're particularly geared to
reducing the memory footprint. Set::Light is supposed to be but I
gather from reading the man page that it's effectively only using
regular hashes to implement set membership. I'm inclined to implement
a solution based on compressing the set contents in memory using Rice
encoding:

 http://giovanni.bajo.it/post/47119962313/golomb-coded-sets-smaller-than-bloom-filters

