-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
I've finished a proof of concept for Metropolis-Hastings corrected
probes in Fred. If you are so inclined, please offer feedback on the
patches. [1] Metropolis-Hastings correction was described previously,
[2] and the probes are exposed over FCP. [3] An originating node starts
a probe request by specifying the type of result they're looking for,
and optionally a value for hopsToLive, which currently defaults to its
maximum of 50 if not specified. We will in practice initially use
something closer to 30. One can disable responding to any number of
result types on the Core settings page.
Possible types are:
BANDWIDTH - outgoing bandwidth limit.
BUILD - Freenet build number. (For instance, currently 1407.) Useful
for measuring update propagation.
IDENTIFIER - endpoint identifier. Useful for improving estimates of
network size and churn.
LINK_LENGTHS - differences in location between endpoint and each of
endpoint's peers.
STORE_SIZE - size of datastore.
UPTIME - session uptime and the percentage of the past 48 hours during
which the endpoint was up.
+/- up to 1% noise is added to bandwidth, link lengths, store size, and
uptime information in an attempt to make the information less
identifiable but still useful for analysis.
Bandwidth, link lengths, store size, and uptime are intended to improve
knowledge of the network so that any problems will hopefully become
apparent, or at the very least allow more detailed simulations which
more accurately reflect actual network conditions. This is with the
goal of implementing changes to improve network performance and
reliability.
Possible outcomes of the probes as implemented are:
-non-propagated timeout
-non-propagated disconnection
-endpoint refusal for the result type
-result from endpoint
Based on feedback in #freenet I've built a TODO list:
1. Access MHProbe from the callbacks, which are inner classes, via
MHProbe.this. This will mean no need to pass it in to use MHProbe as a
ByteCounter. Also because the callbacks are inner classes there is no
need for the horrible hack that is making pendingProbes public static.
2. If FOAF data for a peer is not available, ignore that peer.
3. Implement error messages so that resources can be freed quickly in
the event on an error:
- next in chain disconnected
- timeout
- type unrecognized
- FOAF disabled
4. If FOAF is completely disabled on a node, rather than ignore every
peer, leading to HTL decrementing until a local response, respond with
a "FOAF disabled" error message as above.
5. Add a node configuration option for a random (not based on noderef
like current probe and swap request identifier) identifier which is set
to a random value when at its default of -1.
6. Use a Gaussian distribution for randomNoise() because many tests
assume Gaussian noise, so why not?
7. Quantize things: instead of just adding noise, report bandwidth in
as 64-bit integer KiB and store size as 64-bit integer GiB.
8. Instead of separate identifier and uptime requests, allow requesting:
- [ identifier, quantized noisy 7-day uptime percentage ]
- [ noisy 7-day uptime percentage ]
- [ noisy 48-hour uptime percentage ]
9. Set timeout as constant1 * HTL + constant2 to account for
probabilistic decrement at HTL = 1.
Next week I plan to address the TODO list and any more issues that come
up in the patch set, then submit it for merging into Fred once it has
been approved. After that, I will work on FCP library support for the
new probes, and update pyProbe to support gathering and analyzing the
new information. In the event that I finish these things I can work on
making the network simulator more flexible. Our target deployment date
is Friday the 25th May or the day after.
Thanks,
operhiem1
[1] https://github.com/thynix/fred-staging/tree/newProbes
[2] https://emu.freenetproject.org/pipermail/devl/2012-April/036363.html
[3] https://wiki.freenetproject.org/User:Operhiem1/MHProbeFCP
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)
iQIcBAEBAgAGBQJPuBSaAAoJECLJP19KqmFurkUQAJZbvpgctvsZxgppkRIW/ZOY
meNolQJCutEUHl8VFSdut/6ahRwJ3gfdCgOiF/GHcbEkZqm4S+Jna58Sz6E9WBSa
MfhjQ1dULutiE1vmXN/5v/vNwFYXlATAk2XK35IrUqbCorq7NZHQhSyn61j6fLdB
cGIBOV8GCRHd2rhFu4+oP9tECyoBbL7B9OqjI3AHPEuDEyvFVfuxfk6HDcOam+4Z
dzeecRUMU4T1wquOZ9l6sDTqe8m+RRuR++WHObQ/pI/n1s848HWTM0qs3Dzj3pIU
GNcgX4dbgH8yz3veInur2i6FSkM41IdwEbV3OHro+peOIxVmCJ3c6cEbkEYLQW/R
1mICM7+6CLNkzjrxSwiPNrhVYGGKrYy7F4YSDj2GWwwfwd5aQaxzCM5CO9+VaHWj
IL0b7TvRb2wiNnJWhKlnzNxosZMfQmgGGPskeIImcXUnKPuwy/SfGhEdNifVcsY9
nCF56vGQLmEK/PPdX4CjhWRHO20l2hle5qltapd0VZ9Fyq/pcxw9SiGEW6soBlnw
1gmyxLiWe5PSSic5QD/xtNLWXnPiqnt9gwyaJu0htfZxhB57usrfzivxJGOZbenK
t+QsBGTh8rv0qDHLu1MYKdz0zkGzMSNjNI5dAUs4xKptJ5wXeMsc8K8Ho+Ho7XcF
fxh2AN9VYZSUzTPRAb2L
=rwqn
-----END PGP SIGNATURE-----
