This is ericpony's blog

Thursday, May 28, 2015

A note on formal verification of programs

Formal verification for softwares involves proving properties of programs under various assumptions. In order to automate the verifying process, we have to formalize the program to verify. This usually amounts to transforming the program from source code to model while preserving the properties of interests. In this post, we shall present an example formalization of programs. We shall also demonstrate how program properties can be verified using the presented formalization.

Let's first consider a simple imperative programming language that can encode assignments, arithmetic operations, conditional branches, skips, sequences and non-recursive function calls. Clearly, any program written in this language has a finite number of possible execution paths, as well as a finite number of steps in each execution path. This fact allows us to transform a program to a logical formula, such that each model of the formula corresponds to a valid execution path of the program and vice versa.
More precisely, suppose $F(x_1,...,x_n)$ is a program that takes $n$ inputs and gives one output. Then $F$ can be represented as a formula (called a path formula) in form of $$ \phi_F(x_1,...,x_n,y) = \bigvee\nolimits_{i \in I}\left( p_i \wedge y = e_i \right),$$where $I$ is a finite index set, $p_i$ is a predicate and $e_i$ is an expression. A path formula $\phi$ is called valid if the following conditions are satisfied: i) $\bigvee_{i \in I}\ p_i$ is a tautology, which means all possible paths are covered by the formula, ii) $\neg\bigvee_{i \neq j}(p_i \wedge p_j)$ is a tautology, which means no two paths will be taken simultaneously, and iii) an assignment to $\{x_1,..., x_n, y\}$ satisfies $\phi_F$ if and only if $f(x_1,..., x_n)=y$, which means solutions of the formula precisely characterizes the input-output relation of program $F$.
One can employ techniques of symbolic execution to convert programs of various imperative languages into path formulae, see e.g., [1,2]. Such conversions usually allow one to reduce the problems of verifying program properties to those of checking satisfiability for first-order constraints. For example, to verify that function $F(x_1,x_2)$ is commutative, i.e., $\forall x_1,x_2.\ F(x_1,x_2) = F(x_2,x_1)$, it suffices to check that constraint $\phi_F(x_1,x_2,y) \wedge \phi_F(x_2,x_1,y') \wedge y \neq y'$ is unsatisfiable.
Modern SMT solvers effectively decide formulas over combinations of theories such as uninterpreted functions, boolean algebra, integer arithmetic, bit-vectors, arrays, sets and maps [3]. Program verification thus can benefit from advances in SMT solvers by modelling programs as first-order constraints augmented with theories, see e.g., [4-7].

Handling recursive functions

Uninterpreted functions are often leveraged to facilitate automatic reasoning about recursive functions. The idea is to abstract each recursive call to an uninterpreted function call. If a constraint is unsatisfiable assuming uninterpreted functions, then it is also unsatisfiable assuming the correct interpretation. On the other hand, if a constraint is satisfiable assuming uninterpreted functions, then no reliable conclu-sion could be drawn for the correct interpretation, as the assumptions can be wrong. In [4], the authors proposes a technique to refine assumptions iteratively when the constraint is satisfiable. This technique is the cornerstone of the Leon verification system [5].
As an illustration of the technique, consider function
function size (list) {
  if (list == null) { return 0 } else { return 1 + size(list.next) }
}
with path formula $ \phi_{size} \equiv (b_1 \wedge y = 0) \vee (\neg b_1 \wedge y = 1 + \mathrm{size(list.next)}),$ and define $$ \phi = (\mathrm{size(list)} = y) \wedge (b_1 \Leftrightarrow \mathrm{list} = \mathrm{null}) \wedge \phi_{size}.$$Note that a boolean variable $b_1$, called a control literal, is introduced to represent the truth value of the branching condition in $\mathrm{size}$. The purpose of introducing control literals to a path formula is to guard invocations of uninterpreted functions. It turns out that we can switch between under- and over-approximation of a path formula by changing assignments of its control literals [4], as explained below.
Suppose we want to verify property $\psi$, say $\psi \equiv \mathrm{size(list)} \ge 0$. Define $\psi_1 \equiv \neg\psi \wedge \phi$. Given a correct interpretation of $\mathrm{size}$, $\psi_1$ is satisfiable iff the $\mathrm{size}$ function satisfies property $\psi$. In particular, if $\psi_1$ is unsat, then so is $\psi$ for any interpretation of $\mathrm{size}$. However, if a counterexample is reported assuming uninterpreted functions, then this may be because $\mathrm{size(list.next)}$ is assigned an incorrect value. To exclude such case, we check the satisfiability of $\psi_1 \wedge b_1$. If the later formula is sat, then so is $\psi_1$ and we are done. Otherwise, we have to explore the branch restricted by $b_1$. For this purpose, we unfold the definition of $\mathrm{size(list.next)}$ one more time in $\phi$, and then rewrite formula $\psi_1$ with appropriate substitutions, e.g., replacing $\mathrm{list}$ with $\mathrm{list.next}$, using fresh variables for $b_1$ and $y$, etc. The new formula is denoted as $\psi_2$, and we proceed to check the satisfiabilty of $\psi_2$ and $\psi_2 \wedge b_2$, where $b_2$ is the control literal introduced by the unfolding.
We repeat these steps and produce a sequence of alternating approximations. Thus the verification procedure for property $\psi$ would look like
    $n := 1$ while true do if $\psi_n$ is unsat then output UNSAT and break if $\psi_n\wedge b_n$ is sat then output SAT and break $n := n + 1$ done
Observe that the procedure iteratively computes a succession of under- and over-approximations of the recursive calls in the provided constraint. Also, it finds a counterexample for the constraint whenever the constraint is satisfiable. Intuitively, this happens because a counterexample corresponds to an execution that violates the property, and the procedure enumerates all possible executions in increasing lengths. On the other hand, since our demo language becomes Turing complete after it is equipped with recursions, we cannot expect the procedure to always terminate when the provided constraint is unsatisfiable. For an important class of recursive functions, however, the approach outlined here acts as a decision procedure and terminates in all cases [8]. For instance, the $\mathrm{size}$ function listed above falls into this class.

(To be completed)

References

1. Liu, Chang, et al. "Automating distributed partial aggregation." SoCC, 2014.
2. Srivastava, et al. "Path-based inductive synthesis for program inversion." PLDI, 2010.
3. De Moura, Leonardo, et al. "Satisfiability modulo theories: introduction and applications." Comm. of ACM, 2011.
4. P. Suter, A. S. Koksal, et al. "Satisfiability modulo recursive programs." SAS, 2011.
5. Blanc, Regis, et al. "An overview of the Leon verification system." SCALA, 2013.
6. Bjørner, et al. "Program Verification as Satisfiability Modulo Theories." SMT@ IJCAR, 2012.
7. Bjørner, et al. "Higher-order Program Verification as Satisfiability Modulo Theories with Algebraic Data-types." Higher-Order Program Analysis, 2013.
8. Suter, P., et al. "Decision procedures for algebraic data types with abstractions." POPL, 2010.

Monday, April 6, 2015

Tips for writing Spark programs

Writing serializable closures

When you use a closure, Spark will serialize it and send it around the cluster. This means that any captured variables must be serializable. But sometimes it can't be. Consider this example:
class Foo {
  val factor = 3.14159
  val log = new Log(...) // not serializable
  def multiply(rdd: RDD[Int]) = {
    rdd.map(x => x * factor).reduce(...)
  }
}
The JVM must pass the closure to map to capture variable factor. However, since the closure contains an unserializable variable log, a NotSerializableException will be thrown at runtime. A work around here is to assign factor to a local variable:
class Foo {
  val factor = 3.14159
  val log = new Log(...) // not serializable
  def multiply(rdd: RDD[Int]) = {
    // Only factor2 will be serialized.
    val factor2 = factor 
    rdd.map(x => x * factor2).reduce(...)
  }
}
Serialization is a general issue for distributed programs written on the JVM. A future version of Scala may introduce a "serialization-safe" mechanism for defining closures for this purpose.

Preventing the driver program from OOM

1. Operations on RDDs includes transformations and actions. When the dataset is distributed over the workers, the outputs of transformations are still distributed. In contrast, actions will transfer the outputs to driver program, which can lead to OOM for large outputs. A rule of thumb here is to avoid using actions that may return unbounded outputs, such as collect, countByKey, collectAsMap, etc. If you need to probe a large RDD, try to use actions such as sample, count and take to obtain bounded results. Also, you can write outputs directly from workers to the HDFS using saveAsTextFile and then load them later.
2. Sometimes, you get an OOM Error not because your RDDs don't fit in memory, but because the working set of one of your tasks was too large. Spark’s shuffle operations (sortByKey, groupByKey, reduceByKey, join, etc) build a hash table within each task to perform the grouping, which can often be large. The simplest fix here is to increase the number of tasks, so that each task’s input set is smaller. Spark reuses one executor across many tasks with low launching costs, so you can safely increase the number of tasks to more than the number of cores in your clusters.

Avoiding imbalant join operations

1. Suppose that you want to inner join two RDDs in Spark, one is very small with $n$ partitions and the other is very large with $m\gg n$ partitions. This operation will shuffle the data in the large RDD into $n$ partitions, carried out by at most $n$ executors regardless of the number of work nodes. A better way to do inner join in this case is to first broadcasts the small RDD to the nodes on which the large RDD is distributed, and then performs the join operation at each executor. In this way, you not only avoid the need of shuffling, but also maintain the parallelism of the larger RDD (with $m$ output partitions instead of $n$). The benefits are obtained at the cost of replicating and broadcasting the smaller RDD around the cluster.
2. If you want to left join a very small RDD with a very large RDD, it may be helpful to first filter the the large RDD for only those entries that sharing keys with the small RDD. This filtering step reduces the data needed to be shuffled over the network and thus may speed up the overall execution.

Note that this kind of balancing tricks are not needed for reduce-like operations such as groupByKey and reduceByKey, where the largest parent RDD’s number of partitions will be used.

References and resources

1. The official guidelines for turning Spark
2. How-to: Tune Your Apache Spark Jobs at Cloudera
3. CME 323: Distributed Algorithms and Optimization at Stanford Univ.

Sunday, March 22, 2015

Designing algorithms with Spark aggregate – Part II

In the previous post, I proposed a framework for Spark aggregate and listed several examples derived from the framework. One may observe that the $\mathrm{comb}$ functions are associative and commutative in all but the last example. One may also observe that the results of aggregate calls are deterministic in all but the last example. Now a question rises: is it necessary to use an associative and commutative $\mathrm{comb}$ function in an aggregate call, so that the result can be deterministic? The answer is (somewhat surprisingly) no. In this post, I shall give a simple guideline to find deterministic aggregate examples with non-commutative $\mathrm{comb}$ functions.

Given domain $A$ and range $B$, the idea is to find $B' \subset B$, $\mathrm{seq}: B \rightarrow A \rightarrow B'$ and $\mathrm{comb}: B \rightarrow B \rightarrow B'$, such that $\mathrm{aggregate(zero,\ seq,\ comb,}\ as)$ is deterministic w.r.t. the provided $\mathrm{zero} \in B$ and all $as \in A^+$. Moreover, $\mathrm{comb}$ is non-commutative over $B$, but the restriction of $\mathrm{comb}$ to $B'$ is associative and commutative.

The above idea can be instantiated using an abelian subgroup $(B', ⊗)$ of a non-abelian group $(B, ⊗)$. For example, we can let $B'$ be a cyclic subgroup of a symmetric group $B$. Here ⊗ is function composition and $A$ is the set of integers:

$\quad \mathrm{zero}$: an arbitrary element from $B$.
$\quad \mathrm{seq}\ (b, n) = b^n$.
$\quad \mathrm{comb}\ (b_1, b_2) = b_1 ⊗ b_2$.

Observer that the range of $\mathrm{comb}$ is confined to $B'$ due to the closure property of group operators. For another example, we can let $B$ be the group comprised of all non-singular 2-by-2 square matrices, and $B'$ be the subgroup formed by the rotation matrices in $B$. Here we set $A = B$ and let ⊗ be matrix multiplication:

$\quad \mathrm{zero}$: the identity matrix.
$\quad \mathrm{seq}\ (b, a) = b ⊗ [a]$, where $[x] = x \cdot |\det(x)|^{-1}$, the normalized matrix of $x$.
$\quad \mathrm{comb}\ (b_1, b_2) = b_1 ⊗ b_2$.

We can also construct examples using an abelian base group. We just have to "twist" the $\mathrm{comb}$ function such that it is commutative only over $B'$. For example, let $B$ be the set of n-by-n square matrices, and B' be the subset of $B'$ formed by all matrices whose last row is zero. Moreover, set $A = B$ and use matrix addition $+$ as a commutative group operator. Then we have the following example:

$\quad \mathrm{zero}$: the zero matrix.
$\quad \mathrm{seq}\ (b, a) = b + \{a\}$, where $\{x\}$ is obtained from setting the last row of ${x}$ to zero.
$\quad \mathrm{comb}\ (b_1, b_2) = b_1 + \langle b_2\rangle$, where $\langle x\rangle$ is obtained from doubling the last row of $x$.

Observe that $\mathrm{seq}$ and $\mathrm{comb}$ behave just like ordinary matrix additions over $B'$. They however become non-commutative operators when they are applied to elements outside $B'$. It turns out that we can use a textbook of elementary abstract algebra as a free source of interesting aggregate examples, at least in theory.

Saturday, February 21, 2015

An example of online algorithm

Online algorithms. An online algorithm are designed to receive and serve an unbounded finite sequence of requests $\sigma=\sigma(1),\sigma(2),...,\sigma(m)$ with a bounded memory. The requests are served in the order of occurrence. When serving request $\sigma(t)$, the algorithm doesn't have knowledge of requests $\sigma(t')$ for $t'>t$. The total number of requests is unknown until the last request is served. Serving requests incurs cost. The goal of an algorithm designer is to minimize the total cost paid on serving the entire request sequence.
Competitive analysis. The performance of an online algorithm is usually evaluated by comparing it with an optimal offline algorithm in the worst case. Let $ALG$ be an online algorithm, and $OPT$ be an algorithm that knows all the requests in advance, so that it can serve them in the minimal cost. Given a request sequence $\sigma$, let $ALG(\sigma)$ and $OPT(\sigma)$ denote the costs incurred by $ALG$ and $OPT$, respectively. Algorithm $ALG$ is called c-competitive if there exists a constant $b$ such that for all sequences $\sigma$, $$ALG(\sigma) \le c · OPT(\sigma) + b.$$The competitive ratio of algorithm $ALG$ is defined as $c(ALG)=\sup_{\sigma}ALG(\sigma)/OPT(\sigma).$ The best competitive ratio an algorithm can achieve is thus $c^*=\inf_{ALG}c(ALG)$.

Example. The ski rental problem is one of the classic online optimization problems. The scenario is as follows. You are having a vacation at a ski resort without skis. You can either rent or buy skis from the local shops. You are determined to ski on every sunny day. Suppose the cost of renting skis is one dollar per day, and the cost of buying skis is $s>1$ dollars. Denote the number of sunny days in your vacation as $d$. Clearly, if $d$ is known beforehand (eg. from a magical local weather forecast service), then the optimal (offline) strategy is to rent skis all the way when $d<s$, and buy skis at the first day when $d\ge s$. Now, let's assume you are in the opposite situation, where $d$ is unknown and can't be estimated. You need an online algorithm to help you make decision to rent or buy skis.
Suppose a deterministic online algorithm tells you to rents skis up to day $k$ ($k\le d$) and buy skis on day $k+1$ when $k<d$. Then the competitive ratio of this algorithm is computed by\begin{align*}\label{eq:eg1} \max\left\{\min\limits_{k\ge d}\max\limits_{k\ge d, s\ge d}\left\{\frac{d}{d}\right\}, \min\limits_{k\ge d}\max\limits_{k\ge d, s<d}\left\{\frac{d}{s}\right\}, \min\limits_{k<d}\max\limits_{k< d, s\ge d}\left\{\frac{k+s}{d}\right\}, \min\limits_{k<d}\max\limits_{k < d, s<d}\left\{\frac{k+s}{s}\right\}\right\}. \end{align*}The four terms in the braces correspond to the online/optimal cost ratios in the four cases. In each case, the ratio is first maximized by the adversary (ie. the weather) by choosing $d$ assuming the values of $k$ and $s$. The online algorithm then minimizes the maximal ratio (ie. the worst case) by selecting $k$ based on the choice of $d$. In this way, we obtain a tight upperbound on the cost ratio for each case. Since competitive analysis is a worst-case analysis, the largest bound among the four upperbounds is chosen as the competitive ratio:$$\max\left\{1, \frac{s+1}{s}, \min_{k<s}\left\{\frac{k+s}{k+1}\right\}\right\} = \max\left\{1, \frac{s+1}{s}, \frac{2s-1}{s}\right\} = 2 - \frac{1}{s}.$$This ratio is met by renting for $s-1$ days and buying at the $s$th day when $s\le d.$ No deterministic algorithms can achieve a competitive ratio smaller than $2-\frac{1}{s}$.

Randomized online algorithms. In many problems, online algorithms can achieve a better performance if they are allowed to make random choices. A randomized online algorithm $ALG$ is called $c$-competitive if there are constants $c,b$ such that $E[ALG(\sigma)] \le c ·OPT(\sigma)+b$ for all request sequences $\sigma$, where expectation is taken over the random choices made by $ALG$. The best competitive ratio for randomized online algorithm is thus $c^*=\inf_{ALG}\sup_{\sigma}E[ALG(\sigma)]/OPT(\sigma).$ For simplicity, here we only consider requests generated by an oblivious adversary, that is, an optimal offline algorithm that generates the entire request sequence before any requests are served and any random choices are made. One can also consider competitive ratios against more powerful adversaries; see [3] for more details.

Example. In the ski rental problem, a randomized algorithm chooses some day $k$ at random according to some distribution, rents for up to $k$ days and buys skis at day $k+1$. Assuming the distribution of $k$, the adversary would choose $d$ to maximize the expected online/optimal cost ratio in each case. The actual distribution of $k$ is then determined to minimize the maximal expected ratio over all cases. It turns out that the optimal competitive ratio of randomized online algorithms is roughly $\frac{e}{e-1}\approx 1.58$ [1,2]. (TODO: give the derivation of $\frac{e}{e-1}$.)

Lowerbound analysis. A randomized algorithm may be interpreted as a random choice among deterministic algorithms. In this sense, the problem of finding the best competitive ratio may be viewed as a 2-person zero-sum game, where i) the players are the Algorithm Designer and the Adversary, ii) Adversary’s strategies are all request sequences, iii) Designer's strategies are all deterministic online algorithms, and iv) the payoff for Designer is $ALG(\sigma)/OPT(\sigma)$. In addition, the players are allowed to use randomized strategies. Suppose that Designer chooses algorithms $ALG$ according to probability distribution $P$, and Adversary chooses sequences $\sigma$ according to probability distribution $Q$. Then it follows from Yao's principle that the best competitive ratio $c^*$ for a randomized online algorithm satisfies $$c^*\ge \inf_{ALG}E_Q[ALG(\sigma)/OPT(\sigma)]$$That is, lowerbounds for $c^*$ can be obtained by computing the minimal expected online/optimal cost ratio of deterministic online algorithms against random request sequences. (In fact, using a more general result called von Neumann's Minimax Theorem, one can show that $c^*=\sup_{Q}\inf_{ALG}E_{Q}[ALG(\sigma)/OPT(\sigma)].$)

References

1. https://en.wikipedia.org/wiki/Ski_rental_problem
2. http://cs.brown.edu/~claire/Talks/skirental.pdf
3. http://www14.in.tum.de/personen/albers/papers/inter.pdf

Thursday, February 12, 2015

Designing algorithms with Spark aggregate – Part I

In this post, I shall give a simple framework to design algorithms in terms of Spark's aggregate function. This framework is just a proof of concept and should not be applied blindly in practice. However, it does induce efficient algorithms in some special cases, as will be shown in the examples.

In general, our goal is to compute a property of any given instance from a domain. Let $G$ denote the domain of all interested instances, and $C$ the domain of the desired property. For example, if $G$ denotes graphs and the property to compute is MST, then $C$ will be the set of trees. Furthermore, we assume that there exists an associative and commutative "composition" operator $\oplus : G \rightarrow G \rightarrow G$, such that an instance of interests can be decomposed into (and composed back from) sub-instances in $G$. We shall denote the desired property of instance $g$ as $f(g)$. For example, if $g$ is a graph, then $f(g)$ can be some graph property such as diameter, MST and vertex cover. Note that the the mapping $f : G \rightarrow C$ can be non-deterministic if the property is not unique w.r.t. the given instance.

To use the aggregate function, we specify zero, seg and comb as follows:
$\mathrm{zero} \in (G, C)$ is a pair comprised of the zero values from $G$ and $C$. If there are no such zero values, we can set $\mathrm{zero} = (g, f(g))$ for an arbitrary sub-instance of $g$.
$\mathrm{seq} : (G, C) \rightarrow G \rightarrow (G, C)$ takes a partially aggregated result $(g_1, f(g_1))$ and a sub-instance $g_2$. It outputs $(g_3, f(g_1 \oplus g_2))$ such that $f(g_3 \oplus g) = f(g_1 \oplus g_2 \oplus g)$ for all $g \in G$.
$\mathrm{comb} : (G, C) \rightarrow (G, C) \rightarrow (G, C)$ takes two partially aggregated results $(g_1, f(g_1))$ and $(g_2, f(g_2))$. It outputs $(g_3, f(g_1 \oplus g_2))$ such that $f(g_3 \oplus g) = f(g_1 \oplus g_2 \oplus g)$ for all $g \in G$.
The rest of the steps are standard: we first decompose the input instance into a sequence of sub-instances. The sequence is then converted to an RDD through parallelize. Finally, we invoke RDD.aggregate(zero)(seq,comb) to compute the desired property of the instance.

Remarks

1. While functions seq and comb look very similar in spec, they can use different methods to compute $f(g_3)$. Due to the partial results provided, seq is suitable to use incremental methods and comb can use (the merge part of) divide-and-conquer methods to compute the desired property. See the convex hull example below.
2. The assumption that operator $\oplus$ is associative and commutative over $G$ can be relaxed to be so only with respect to the given property $f$. For instance, suppose that $G$ is a set of finite strings and $\oplus$ is the string concatenation operator. While $\oplus$ is not commutative per se, it is commutative w.r.t. the string length property. That is, $\mathrm{strlen}(g_1 \oplus g_2) = \mathrm{strlen}(g_2 \oplus g_1)$ for all $g_1, g_2 \in G$. The framework is thus applicable to compute the length of string. See the majority algorithm below for a more non-trivial example.
3. The framework is practical only if $|g_3|$ can be significantly less than $|g_1 \oplus g_2|$, for otherwise we will have to pass almost the whole instance from node to node. In the case that $f(g_1 \oplus g_2)$ can be computed from $f(g_1)$ and $f(g_2)$, there is no need to output $g_3$, i.e., $|g_3|=0$. This special case allows us to operate on simpler domains (e.g., the type of seq becomes $C \rightarrow G \rightarrow C$) and to obtain more efficient aggregate algorithms due to less communication overheads.

Examples

The above naive framework can yield efficient algorithms if $f(g_1 \oplus g_2)$ can be computed solely based on $f(g_1)$ and $f(g_2)$. Below we give three examples that satisfy this condition. The first is a variant of the reverse-delete algorithm that finds an MST of a graph. The second is an outline of a convex hull algorithm. The last is Boyer and Moore's algorithm that finds the majority element in a list.

Finding MST. Let $G$ be the set of all weighted undirected graphs, and $C$ be the set of trees. Given a graph $g$ from $G$, we assign a unique index to each of its edges. Let $w(e)$ and $idx(e)$ denote the weight and index of edge $e$, respectively. Given two edges $e_1$ and $e_2$ of $g$, we say $e_1 < e_2$ iff either $w(e_1) < w(e_2)$, or $w(e_1) = w(e_2)$ and $idx(e_1)$ < $idx(e_2)$. Now we define an operator msf over $G$, such that msf computes a minimum spanning forest of $g_1 \cup g_2$ from $g_1$ and $g_2$ as follows:
def msf (g1, g2) = {
  var g = g1 ∪ g2
  while (g contains a cycle c)
    remove the largest edge from c
  return g
}
Let RDD(g) denote the RDD obtained from decomposing and parallelizing graph $g$. Then RDD(g).aggregate(Nil)(msf,msf) computes a minimum spanning forest of $g$, which would be a MST if $g$ is connected. Note that the result is deterministic if the assignment of indices is deterministic.

Finding convex hull. Here $G$ is the collection of finite sets in Euclidean plane and $C$ is the set of convex polygons. The convex hull of a set of points is the smallest convex polygon containing all the points. If the points are given as a sequence, we can use aggregate to find their convex hull by computing smaller convex hulls incrementally in seq and then merging them in comb to obtain the final convex hull. Both of seq and comb can be implemented to take linear time (see the links for pseudo codes).

Finding majority. Let $C$ be a set of comparable elements. Fix some element $\bot \in C$ and let $G = (C - \{\bot\})^*$ be a set of finite sequences. The majority problem is to determine in a given sequence from $G$ whether there is an element with more occurrences than all the others, and if so, to determine this element. Click here to see a linear-time algorithm that solves the problem using aggregate. Note that the aggregate invocation only finds a candidate for majority; we need a further check to assert that the candidate is indeed the majority element.

Friday, January 23, 2015

Deterministic aggregate semantics in Scala and Spark

The aggregate function in Scala has method signature as follows:
def aggregate[A](z: => A)(seqOp: (A, B) => A, combOp: (A, A) => A): A
The semantics of aggregate depend on the collection upon which it is invoked. When aggregate is invoked on a sequential traversable collection such as a List, it fallbacks to foldLeft and ignores the provided compOp. In this case, aggregate behaves just like the ordinary fold in typical functional programming languages such as Lisp and ML. When aggregate is invoked on a parallel traversable collection, however, it performs a two-stage aggregation over the data elements in parallel. In Scala, a parallel collection can be splitted into partitions and distributed over several cores or nodes. Invoking aggregate on a parallel collection reduces a value for the collection by first folding each partition separately using seqOp, and then merging the results one by one using combOp.

Note that the Scala API document doesn't specify the way in which the collection is partitioned, as well as the order in which the intermediate results are merged. Hence, if a programmer hopes to obtain the same output whenever aggregate is invoked on the same parallel collection, he has to make sure himself that seqOp is associative and combOp is both associative and commutative.

Aggregate v.s fold

The fold function in Scala has method signature as follows:
def fold[A1 >: A](z: A1)(combOp: (A1, A1) => A1): A1
One can observe two differences between the signatures of fold and aggregate: 1) fold only needs one operator, while aggregate needs two; 2) fold requires the result type to be a supertype of the element type, while aggregate doesn't. Note that both methods fallback to foldLeft when they are invoked on a sequential collection. One may notice that Scala's aggregate is in effect the ordinary FP fold
fold :  (A → B → A) → A → ([B] → A)
and Scala's fold is just a special form of its aggregate. The idea behind this design is to allow for parallel FP fold. In Scala, fold and aggregate are both done in two stages, namely folding and combining. For fold, the same operator is used for folding and combining, and this requires the supertype relationship between the element and the result type. Without the supertype relationship, it will take two operators to fold and combine, and this is why Scala provides aggregate that takes two operators as arguments.

Aggregate in Spark

Spark implements its own aggregate operations for RDD. One can observe from the source code of Spark that the intermediate results of aggregate are always iterated and merged in a fixed order (see the Java, Python and Scala implementations for details). Hence, we don't have to use commutative combOp to obtain deterministic results from aggregate, given that at least one of the following conditions holds:
(i) there doesn't exist an empty partition in any partitioning of the collection, or
(ii) z is an identity of combOp, i.e., for all a of type A, combOp(z,a) = combOp(a,z) = a.
Note that this observation is not assured by the official Spark API document and is thus not guaranteed to hold in the future versions of Spark. However, it is still useful if we want to model the behaviours of a Spark program formally.

Below, we shall try to formalize Spark's implementation of aggregate. While Scala is an impure functional language, we only consider pure functions here for simplicity. Let $R(L,m)$ denote an RDD object obtained from partitioning a list $L \in \mathbb{D}^*$ into $m \ge 1$ sub-lists. Suppose that we have fixed an element $zero \in \mathbb{E}$, as well as operators $seq:{\mathbb{E} \times \mathbb{D} \to \mathbb{E}}$ and $comb:{\mathbb{E} \times \mathbb{E} \to \mathbb{E}}$. An invocation of aggregate is said to be deterministic with respect to $zero$, $seq$ and $comb$ if the output of $$R(L,\cdot).\textrm{aggregate}(zero)(seq,\ comb)$$ only depends on $L$. Define a function $\textrm{foldl}:({\mathbb{E} \times \mathbb{D} \to \mathbb{E}}) \times \mathbb{E} \times \mathbb{D}^* \to \mathbb{E}$ by \begin{eqnarray*} \textrm{foldl}(f,\ a,\ Nil) & = & a \\ \textrm{foldl}(f,\ a,\ b \!::\! bs) & = & \textrm{foldl}(f,\ f(a,b),\ bs), \end{eqnarray*} where $f\in {\mathbb{E} \times \mathbb{D} \to \mathbb{E}}$, $a \in \mathbb{E}$, $b \in \mathbb{D}$ and $bs \in \mathbb{D}^*$. Suppose that $1\le m \le |L|$ and $R(L,m)$ partitions $L$ into $\{L_1, \dots, L_m\}$ such that $L=L_1 \cdots L_m$. According to the source code of Spark, we can define the aggregate operation in Spark as \begin{eqnarray*} R(L,m).\textrm{aggregate}(zero)(seq,\ comb) = \textrm{foldl}(comb,\ zero,\ L'), \end{eqnarray*} where $L' \in \mathbb{E}^*$ is a list of length $m$ and $L'[i] = \textrm{foldl}(seq,\ zero,\ L_i)$ for $i=1,\dots,m$.

Note that the partitioning of $L$ is unspecified and may be non-deterministic. Hence, we have to use an associative $comb$ operator if we want to ensure that the result of aggregation is deterministic. Suppose that $comb$ is associative and at least one of conditions (i) and (ii) is satisfied. It turns out that the output of aggregate is deterministic with respect to $zero$, $seq$ and $comb$ if and only if for all lists $L_1$ and $L_2$, $$comb(zero,\ \textrm{foldl}(seq,\ zero,\ L_1L_2)) = comb(\textrm{foldl}(seq,\ zero,\ L_1),\ \textrm{foldl}(seq,\ zero,\ L_2)).$$ While the above condition is undecidable in general, it can be effectively checked for some classes of programs such as FSMs over finite alphabets.

Concluding remarks

Aggregate in Spark is the parallelized version of the conventional fold, which was already shown to be extremely useful in functional programming. It is therefore of practical interests to study the properties and behaviours of the aggregate function. In particular, it is interesting to establish and verify the conditions under which aggregate is deterministic with respect to the provided $zero$, $seq$ and $comb$. We may further investigate this problem in the coming posts.

Related Posts Plugin for WordPress, Blogger...