Questions and discussions related to the FRESCO project (https://github.com/aicis/fresco)
One thing that comes to mind is reducing the modulus bit length and max bit length to 128 and 64 respectively. That should noticeably speed up your overall application. It may or may not give you enough precision (depending on your concrete computation). If you start seeing incorrect outputs as a result it means that your modulus does have to be bigger. The modulus bit length and max bit length are set on the
SpdzResourcePool class.
Ah thanks for the input, I am actually already setting those values over command line trying to figure out what the best configurations are!
Another question then came up: Does decreasing the mod bit length impact the overall security of the protocol or is it just some internal setting?
Another question then came up: Does decreasing the mod bit length impact the overall security of the protocol or is it just some internal setting?
Yep, it does also affect security. If you are using comparisons (or other computations based on bit decomposition), you get 64 for the computational security parameter and 60 (need to double-check that) for the stat. security parameter when you set mod bit length to 128. These are accepted as sufficiently large in the current MPC literature. If you go lower with the mod. bit length however your computational and statistical security params also decrease (because their sum is >= mod. bit length), so things get controversial.
I totally support that point
@rw0x0 hey! sorry for the late reply. there isn't an "out-of-the-box" way to measure offline and online times separately, however there are two things you could do: 1) run the application you want to benchmark with mascot pre-processing and then with dummy pre-processing. the difference in runtime will roughly correspond to the overhead introduced by mascot. 2) you could figure out how much pre-processing material (e.g., multiplication triples) your application needs and measure the time it takes to generate the required amount with Mascot by using the Mascot class (https://github.com/aicis/fresco/blob/master/tools/mascot/src/main/java/dk/alexandra/fresco/tools/mascot/Mascot.java#L27) directly. hope this helps!
hi, thank you for your reply. I thought that I would need some kind of workaround, what i did in the end was, that i added some functions to the spdzmascotdatasupplier, which allowed me to produce a number of triples (and randomelements, etc.) beforehand.
Something like:
public void produceTriples(int el) {
ensureInitialized();
triples.addAll(mascot.getTriples(el));
}
Calling this function before the actual computation had the result, that during the online phase no triples had to be produced. Maybe something like this could be added to the main project? :)
Hi, just wanted to let you know: I was asked by Marcella from the MPC-SoK project to add the FRESCO sample programs I wrote last year to the MPC-SoK repo. Which I did in this PR MPC-SoK/frameworks#21
I hope that is fine with everyone. I didn't really do much to validate that the old code is working an so on, and it might be out of date with recent FRESCO versions (it uses 1.1.2). If she has problems with it I hope you can help her out.
Is there built in functionality to help calculate the communication costs involved while executing a FRESCO app?
Hi, you might want to take a look at the networkloggingdecorator class which can be found in core/logging.
Using this in addition to the socket network tracks all network traffic one party receives over the network is tracked. This can then be analyzed using the getLoggedValues function.
Using this in addition to the socket network tracks all network traffic one party receives over the network is tracked. This can then be analyzed using the getLoggedValues function.
Hi, is there an easy way to run the examples in https://github.com/aicis/fresco-samples with proper preprocessing?
Oh no! The logo is no longer available at https://alexandra.dk/logo/fresco-logo-artwork.svg ... this messes up the README.md! Someone save the logo!
Bumping the version in https://github.com/aicis/fresco-samples gives loads of errors. Are there any plans to update the examples?
@mkskeller Sure -- let me just check how much work it is.
because of this, you can no longer access library functions directly through the ProtocolBuilder, but need to create ComputationDirectories as you need them. This is done using the syntax "AdvancedNumeric.using(builder)"
@jonas-lj Is there a good way too profile a demo application? I am trying to get a good overview on the DistanceDemo in what methods how much time is used. I want to profile on a different machine without gui, so sadly I can't use any gui profilers. (Otherwise the Intellij Profiler worked quite well for me). Any ideas?
@pffrandsen What do you mean with detailing the testscenario more? I understand that MPC is different then a normal application, nonetheless, I want to capture a complete run to compare it with a complete run after a certain commit. If I run all the nodes locally the network wait should be practically 0?
The question is what do you want to learn/understand. Running all the nodes on a local network might be a constructed example. My go-to benchmark scenario is "Assume all nodes are in the same metro area, say Chicago, Frankfurt or Copenhagen". This is not unreasonable requirement - I have actually been presented to this by a customer. Now we can derive a set of constraints: 100MBit/s, small packages, drop rate of less than 1%, HW can be any machine -> 20 core 64 GB suffices.
If you measure where the bottlenecks are in a local networksetup -> be aware that this does not have any correlation to a real world application,
I am aware that this will be a constructed example, and eventually I will also measure it with some network constraints, like you just described. But for the moment I want to focus on the computational side. The focus atm is OT. I guess I will either try to profile the Demo, or maybe the OT (base) tests and tweak them a bit to be more realistic.