# The variable speed of SAT solving

Vegard Nossum asked me about the varying time it took for CryptoMiniSat to solve a certain instance that was satisfiable. This inspired me to write an overly long reply, which I think might interest others. So, why does the solving time vary for a specific instance if we permutate the clauses? Intuitively, just by permutating the clauses, the problem itself doesn’t get any easier or harder, so why should that make any difference at all? I will now try to go through the reasons, though they can almost all be summed up as follows: the SAT solver has absolutely no clue what it is solving, and so to solve a problem, it relies on heuristics and randomisation, both of which are influenced by its starting seed, which in turn is influenced by the order of the clauses in the CNF file. And now on to the long explanation.

Let’s divide problems into two categories: satisfiable and unsatisfiable instances. First, let me talk about satisfiable instances. Let’s suppose that we have been solving the problem for some time, and we have connected the dots (variables) well enough through learnt clauses. All that the SAT solver is waiting for is a good guess to set 3-4 variables right, and then it will propagate all variables to the right setting to satisfy all clauses. The problem is therefore to get to the point of only needing to guess 3-4 variables, i.e. to connect the dots, which we do through resolution. We can sacrifice some or all of dot-connecting (i.e. resolution) with some more guessing, and this can in fact be coded down by simply foregoing some of the conflict analysis we do. In the extreme case, all conflict analysis can be disabled, and then the solver would not restart its search. The solver would in essence be brute-forcing the instance through BCP (boolean constraint propagation). It is easy to see why this latter is prone to variation: depending on the ordering of the variables in the search tree, the tree could be orders of magnitude smaller or larger, and the first solution can be at any point in the search tree, essentially at a random place.

If we decide to carry out resolution for satisfiable problems to counter the problem of the variation of the search-tree, it is interesting to realise the following: in most cases, we can not forgo guessing. The reason is simple yet leads to quite interesting properties. Essentially, a SAT instance can have multiple solutions. If there are two solutions, e.g. `01100...` and `10011...` i.e. the solutions are the complete inverse of one another, then the SAT solver will not be able to prove any variable to any value. The best it could do is to create binary clauses, in the example case for instance

`var1=0 -> var2=1, var3=1, var4=0...`
and
`var1=1 -> var2=0, var3=0, var4=1...`

If we do enough resolutions, the “guessing” part will eventually become a solution-selector, i.e. it will select a solution from the set of available solutions. If there are 2^20, evenly distributed in the search space, we might need to set 20 variables before a solution is found through a simple application of BCP. Naturally, SAT solvers don’t normally do this, as they are content in finding one satisfying assignment, reporting that to the user, and exiting. It would be interesting to know the average remaining number of variables that needed to be guessed to solve the solution at the point the SAT solver actually found the solution, but this has not been done yet as far as I know. Then, we would know the trade-off the SAT solver employs between resolution and searching when solving satisfiable instances.

All right, so much for satisfiable instances. What happens with UNSAT instances? Well, the solver must either go through the whole tree and realise there is no solution, or do resolution until it reaches an empty clause, essentially building a resolution tree with an empty clause at its root. Since both of these can be done at the same time, there is a similar trade-off as above, but this time it’s somewhat upside-down. First of all, the search tree can be smaller or larger depending on the ordering of the variables (as before), and secondly, the resolution tree can be smaller or larger depending on the order of resolutions. The minimal resolution tree is (I believe) NP-hard to find, which doesn’t help, but there is at least a minimum resolution tree that limits us, and there is a minimum search tree which we must go through completely, that limits us. So, in contrast to finding satisfying solutions, both of these are complete in some sense, which should make searching for them robust in terms of speed. Finding a satisfying solution is not complete, because, as I noted above, SAT solvers don’t find all satisfying solutions — if they did, they would actually have the same properties as solving an unsatisfiable problem, as they would have to prove that there are no more solutions remaining, essentially proving unsatisfiability.

The above reasoning is probably the reason why SAT solvers tend to behave more robustly when they encounter an unsatisfiable problem. In particular, CryptoMiniSat’s running time seems more robust when solving unsatisfiable problems. Also, for problems that contain a lot of randomly placed solutions in the search space, such as the (in)famous vmpc_33 and vmpc_34 problems, the solving times seem wholly unpredictable for at least two interesting SAT solvers: lingeling and CryptoMiniSat.

# Optimisations, take two

I have talked about optimisations before and I mentioned that there are a few left to describe. In this post I will talk about two of them.

The first optimisation I would like to talk about concerns compilation flags. Somehow, everyone seems obsessed with them, like if they could somehow make a slow program fast. Alas, this is not the case. Firstly, and this is quite funny: the gcc flag “-O3” usually turns on all extra flags that people tend to give. One should really ponder upon this when looking at desperate attempts to speed up code. However, there is exactly one flag-combo that is very-very useful besides “-O3”: it’s “-fprofile-generate” and “-fprofile-use”. To understand why these are useful, we must first understand that one of the challenges faced by an optimising compiler is to try to guess how many times a loop will be executed,and how many times a branch will be taken. Given a good set of guesses, the loop can be unwound (or not) and the branches can be taken by default (or not). If compiled with “-fprofile-generate”, the program generates such information on-the-fly, which later can be used by “-fprofile-use”. The speedup achieved with such an approach in the realms of DPLL-based SAT solvers is relatively high, in the order of ~5-10%. I believe many SAT solvers’ binaries don’t use this trick, even though it’s cheap to use. I personally compile with “-fprofile-generate”, then run an example problem like UTI-20-10p1 on it, and then recompile with “-fprofile-use”. An important aspect, by the way, is to execute your problem with a very typical scenario: strange or non-normal scenarios will produce spurious branching and loop data which will actually slow down the program most of the time instead of speeding it up.

The second optimisation I would like to talk about concerns cache usage. Modern processors are extremely fast, to the point that many times what really limits the processor is no longer the ability to execute the code on the data, but to actually get the data that the code needs. This problem is usually alleviated through spatial and temporal data locality that is, that data most likely needed next is usually physically close to the one that we just accessed. However, in DPLL-based SAT solvers, the main function, propagate() goes through a list of pointers, accessing the data where each pointer points to. In other words, the CPU will fetch the next pointer in the background, but will not fetch where that pointer points to — simply because it doesn’t know it’s a pointer. So, how could we alleviate this problem? Well, by telling the CPU that the next piece of information that will be needed is where the next pointer points to. Like this:

```for (i = 0; i < pointers.size();  i++) {
if (i+1 < pointers.size())
__builtin_prefetch(pointers[i+1]);
treat_data(*pointers[i]);
}```

where `pointers` is a vector of pointers that must each be treated. The function `__builtin_prefetch()` calls an assembly instruction that non-blockingly fetches the data pointed to into the Level 1 cache of the processor. This/these instruction(s) are not implemented in all CPUs, but many x86 and all AMD64 architectures have them. Essentially, while we are dealing with data where `pointers[i]` points at, we fetch the data where `pointers[i+1]` points at, in parallel. Interleaving the fetching of next data with the treatment of current data helps us do more things in less time, speeding up the program. Interestingly, since prefetching the data does not alter the flow of the program in any way, omitting the prefetch call does not alter the results of the program at all: it just makes it run slower. In CryptoMiniSat I use this optimisation, and it seems to speed up the solving by a couple of percentages, maybe 5-8%. The amount of time you will save on this depends on the locality and size of the information your pointers are pointing to. If they are very small, e.g. 2-3MB in total and are close to each other, then the CPU will fetch these after the first couple of pointers have been dereferenced, so there is no need to do explicit prefetching. But if the data you are dealing with is much larger, which is often the case with SAT solvers, then such explicit pre-fetching can save quite some time.

# On failed literal probing

Apparently, there have been quite some work done on failed literal probing, although I don’t think they have all been put into a common package yet.

The idea is quite simple in its purest form: try to set variable `v1` to `true` and see if that fails. If it does, variable `v1` must be `false`. If it doesn’t fail, try the other way around: set it to `false` and if that fails, it must be `true`.

There are a couple of tricks we can add, however. For example, if both `v1` and `!v1` set variable `v10` to value `X` (where `X` can be either `true` or `false`), then set `v10` to `X`. The reasoning here is simple: `v1` must be set, so whatever both `v1` and `!v1` imply, it must be set, too. So we can safely set `v10` to `X`.

A more interesting thinking is the following. If `v1` sets `v20` to `true`, but `!v1` sets `v20` to `false`, then `v1 = v20`. So, we can replace `v20` with `v1`. One less variable to worry about!

There are even more tricks, however. If setting `v1` to `true` and `false` both shorten a longer XOR to a 2-long XOR “`v40 + v50 = false`“, this 2-long XOR can be learnt: `v40` can be replaced with `v50`.

And more tricks. If there is a 2-long clause `v1 or v2 = true`, then we can do all of the above, but with `v1` and `v2` this time. Since either `v1` or `v2` must be `true`, all the above ideas still work. In all previous ideas all we used was the fact that either `v1` or `!v1` must be true. This is still the case: either `v1` or `v2` must be true. This, by the way, is called 1-recursive learning. (note: the fourth paragraphs changes a bit, but we still learn a 2-long xor).

And now something new. I have been thinking about trying to re-use old learnt clauses. They really code down sometimes quite important facts about the problem. What if we tried to re-use them? After all, keeping them in memory is really cheap: the only problem with them is that doing regular propagation with them takes a lot of time, thus slowing down the solver. But what if we only used them to do failed literal probing? I just launched a test on the Grid’5000 cluster to find out. Fingers crossed…