On the Complexity of Minimizing Energy Consumption of Partitioning DAG Tasks

Abstract

We study a graph partition problem where we are given a directed acyclic graph (DAG) whose vertices and arcs can be respectively regarded as tasks and dependencies among tasks. The objective of the problem is to minimize the total energy consumed for completing these tasks by assigning the tasks to $k$ heterogeneous machines. We first show that the problem is NP-hard. Then, we present polynomial-time algorithms for two special cases where there are only two machines and where the input DAG is a directed path. Finally, we study a natural variant where there are only two machines with one of them being capable of executing a limited number of tasks. We show that this special case remains computationally hard.

[lc]organization=Department of Computer Science, TU Dortmund University,city=Dortmund, postcode=44227, country=Germany [yyj]organization=Chair of Economic Theory, Saarland University, city=Saarbrücken, postcode=66123, country=Germany

1 Introduction

Tasks that are represented by directed acyclic graphs (DAG tasks) are ubiquitous in many applications, including for instance cloud computing, deep neural network, workflow scheduling, etc. [1, 2, 3, 4, 5, 6]. In this paper, we investigate the complexity of a new graph partition problem which models the scenario where DAG tasks are deemed to be assigned to $k$ heterogeneous machines (e.g., execution units in distributed systems, clusters of cores in heterogeneous multicore systems, etc.), with the objective to minimize the energy consumption for the computation of these tasks under natural restrictions. More precisely, in this problem, we are given a DAG whose vertices represent tasks, in which the energy consumption of a task depends on which machine is allocated for its execution. An arc from a task $a$ to a task $b$ means that the computation of $b$ requires the output of task $a$. So, if $a$ and $b$ are assigned to different machines, the output of $a$ needs to be transferred to the machine executing the task $b$, which also incurs energy consumption. We note that when there are multiple outneighbors of a task $a$ assigned to a machine $i$ different from that of $a$, we need only to transfer once the output of $a$ to the machine $i$.

We assume that the energy consumption associated with the transfer primarily depends on the volume of transferred data and the impact of other factors is neglectable. Under this assumption, we may simply use a univariate function map** tasks to numbers to define energy consumption for data transformation. A more general case is when, in addition to the volume of the data, energy consumption also takes into account the types of data, the identities of machines between which the data is transferred, etc. We prove that even in our special case, the problem is already NP-hard.

Now we formulate the problems. We assume the reader is familiar with the basics in graph theory [7] and parameterized complexity theory [8, 9, 10].

A directed graph (digraph) is a tuple $G=(V,A)$ where $V$ is a set of vertices and $A$ is a set of arcs. Each arc is defined as an ordered pair of vertices. An arc from a vertex $v$ to a vertex $u$ is denoted by $(v,u)$. We say that the arc $(v,u)$ leaves $v$ and enters $u$. We also use $A(G)$ to denote the set of arcs of $G$, and use $V(G)$ to denote the set of vertices of $G$. The set of outneighbors (respectively, inneighbors) of a vertex $v\in V(G)$ is defined as $\Gamma^{+}_{G}{(v)}=\{u\in V(G):(v,u)\in A(G)\}$ (respectively, $\Gamma^{-}_{G}{(v)}=\{u\in V(G):(u,v)\in A(G)\}$). The outdegree (respectively, indegree) of $v$ is the number of outneighbors (respectively, inneighbors) of $v$ in $G$. The set of closed outneighbors (respectively, closed inneighbors) of $v\in V(G)$ is defined as $\Gamma^{+}_{G}{[v]}=\Gamma^{+}_{G}{(v)}\cup\{v\}$ (respectively, $\Gamma^{-}_{G}{[v]}=\Gamma^{-}_{G}{(v)}\cup\{v\}$). Vertices of outdegree zero are called sinks, and those of indegree zero are called sources of $G$. A DAG is a digraph without directed cycles.

For $v\in V(G)$ and $S\subseteq V(G)$, let $\Gamma^{+}_{G}{(v,S)}=\Gamma^{+}_{G}{(v)}\cap S$. In addition, we define

$${\bf{1}}_{G}(v,S)=\begin{cases}1,&\text{if }\Gamma^{+}_{G}{(v,S)}\neq\emptyset,\\ 0,&\text{otherwise}.\\ \end{cases}$$

For an integer $i$, we use $[i]$ to denote the set of all positive integers no greater than $i$. For a function $f:X\rightarrow Y$, we use $f^{-1}(y)$ to denote the set consisting of all $x\in X$ such that $f(x)=y$. For two functions $f:V(G)\rightarrow[k]$ and $q:V(G)\rightarrow\mathbb{R}_{\geq 0}$, we define

$${\textsf{Tran}_{G}(f,q)}=\sum_{i\in[k]}\sum_{v\in f^{-1}(i)}\sum_{j\in[k]\setminus\{i\}}q(v)\cdot{\bf{1}}_{G}(v,f^{-1}(j)).$$ (1)

In Equality (1), $f$ is a function assigning vertices of $G$ to $k$ machines, $q$ defines the energy consumption for transferring the outputs of vertices in $G$, and ${\textsf{Tran}_{G}(f,q)}$ is exactly the total amount of energy consumption for transferring all necessary data under the assignment $f$.

Energy-Saving Partition of DAG (ESP-DAG)
Input:	A DAG $G$, two functions $p:V(G)\times[k]\rightarrow\mathbb{R}_{\geq 0}$ and $q:V(G)\rightarrow\mathbb{R}_{\geq 0}$, a rational number $r$.
Question:	Is there an assignment function $f:V(G)\rightarrow[k]$ so that $$\left(\sum_{v\in V(G)}p(v,f(v))\right)+{\textsf{Tran}_{G}(f,q)}\leq r?$$

In the formulation, $p(v,i)$ for each $v\in V(G)$ and each $i\in[k]$ is the amount of energy consumed for the computation of task $v$ in the machine $i$.

We also study a natural variant of the special case where $k=2$ under the restriction that one of the two machines is capable of executing a limited number of tasks. For simplicity of presentation, we reformulate the special case as follows. For a graph $G$, two disjoint subsets $V_{1},V_{2}\subseteq V(G)$, and a function $q:V(G)\rightarrow\mathbb{R}$, let

$${\textsf{BTran}_{G}((V_{1},V_{2}),q)}=\sum_{i\in[2]}\sum_{v\in V_{i}}q(v)\cdot{\bf{1}}_{G}(v,V_{3-i}).$$ (2)

Size Bounded Energy-Saving Bipartition of DAG (SB-ESBP-DAG)
Input:	A DAG $G$, two functions $p:V(G)\times[2]\rightarrow\mathbb{R}_{\geq 0}$ and $q:V(G)\rightarrow\mathbb{R}_{\geq 0}$, a rational number $r$, and an integer $\ell$.
Question:	Are there disjoint $V_{1},V_{2}\subseteq V(G)$ such that $V_{1}\cup V_{2}=V(G)$, $|V_{1}|\leq\ell$, and $$\left(\sum_{v\in V_{i},i\in[2]}p(v,i)\right)+{\textsf{BTran}_{G}((V_{1},V_{2}),q)}\leq r?$$

Our Main Contributions

We first establish a complexity dichotomy for ESP-DAG with respect to the number of machines: the problem is NP-hard if there are at least three machines (Theorem 1), but becomes polynomial-time solvable when there are two machines (Theorem 2). Afterwards, we show that when the input DAG degenerates to a directed path, the problem admits a pseudo polynomial-time algorithm, regardless of the number of machines (Theorem 3). In contrast to ESP-DAG, we show that the size-bounded variant SB-ESBP-DAG, where there are only two machines, is computationally hard: it is W[1]-hard with respect to the parameter $\ell$ (Corollary 3). As a byproduct of this result, we show that a variant of the minimum cut problem is W[1]-hard with respect to a natural parameter (Theorem 4), strengthening its NP-hardness studied in [11].

2 Related Works

To the best of our knowledge, ESP-DAG has not been studied from a theoretical point of view in the literature. However, the problem and its variant defined above belong to the gigantic family of graph partition problems which have been extensively and intensively studied in the literature. These problems aim at dividing either the vertices or the arcs of a given graph (undirected or directed) into several sets so that certain structural properties are met or certain optimization objections are achieved (see, e.g., [12, 13, 14, 15, 16, 17, 18, 19]).

More specifically, our problem falls into the category of task allocation problems [20, 21, 22, 23]. In the realm of task allocation problems, partitioning a DAG can achieve various objectives. These objectives include minimizing makespan, reducing energy consumption, decreasing communication costs, achieving load balancing, ensuring fault tolerance, etc. Our model aims to minimize energy consumption. In most of the previous works tackling DAG tasks, the cost of data transformation are arc-wisely defined: if a task $v$ assigned to a machine $i$ has multiple outneighbors assigned to a different machine $j$, the output of $v$ needs to be transferred multiple times from machine $i$ to machine $j$, one for each of $v$’s outneighbors assigned to the machine $j$. We build upon previous approaches by considering a more comprehensive energy consumption model for data transformation. We simplify multiple transfers of output data from one machine to another when a task has multiple outneighbors assigned to different machines. This refined modeling approach enables us to evaluate and minimize energy consumption more accurately in DAG-based task allocation.

One noticeable related problem where a similar energy model as ours is adopted is a one studied by Hu et al. [24] in 2019. Particularly, this problem takes as the same input as SB-ESBP-DAG with $\ell$ being dropped, and the problem consists in dividing $V(G)$ into two sets $V_{1}$ and $V_{2}$ to minimize $\sum_{v\in V_{i},i\in[2]}p(v,i)+\sum_{v\in V_{1}}q(v)\cdot{\bf{1}}_{G}(v,V_{2})$, under the restriction that there are no arcs from $V_{2}$ to $V_{1}$. For the problem, Hu et al. [24] proposed a polynomial-time algorithm which is, however, pointed out to be flawed by Li et al. [25].

Our studied problem is also related to several specific resource allocation problems. Particularly, when the given DAG does not contain any arc (or $q(v)=0$ for all $v\in V(G)$), ESP-DAG is equivalent to the problem of maximizing social welfare in multi-agent resource allocation when agents hold $1$-additive utility functions, which is known to be polynomial-time solvable [26, 27]. This special case is also related to a winners allocation problem proposed by Yang [28], which generalizes the multi-agent resource allocation with two agents, with each holding a $1$-additive utility function.

3 Problems to Establish Our Results

Our results are obtained based on the following problems.

An undirected graph $G=(V,E)$ is a tuple where $V$ is a set of vertices and $E$ is a set of edges. The set of vertices and the set of edges of $G$ are also denoted by $V(G)$ and $E(G)$, respectively. An edge between two vertices $v$ and $u$ is denoted by $\{v,u\}$.

For a function $w:S\rightarrow\mathbb{R}$ and a subset $S^{\prime}\subseteq S$, we define $w(S^{\prime})=\sum_{s\in S^{\prime}}w(s)$. For a graph $G$ and a subset $E^{\prime}\subseteq E(G)$, we use $G-E^{\prime}$ to denote the graph obtained from $G$ by removing all edges in $E^{\prime}$.

Multiway Cut
Input:	An undirected graph $G$, a weight function $w:E(G)\rightarrow\mathbb{R}_{\geq 0}$, a set $T$ of $k$ distinct vertices from $V(G)$ called terminals, and a number $r$.
Question:	Is there a subset $E^{\prime}\subseteq E(G)$ such that $w(E^{\prime})\leq r$ and vertices from $T$ are pairwise disconnected in $G-E^{\prime}$?

Equivalently, Multiway Cut determines if we can partition $V(G)$ into $k$ sets, each containing exactly one terminal, so that the total weight of edges crossing the partition does not exceed $r$. It is known that Multiway Cut is NP-hard for every $k\geq 3$, but becomes polynomial-time solvable if $k=2$ [29]. (Multiway Cut with two terminals is exactly the decision version of the classic problem Minimum $s$-$t$-Cut.)

The problem Size Bounded Minimum $s$-$t$-Cut (SBM-$s$-$t$-Cut) takes as the same input as Multiway Cut where $k=2$ together with an integer $\ell$, and determines if there is a bipartition $(V_{s},V_{t})$ of $V(G)$ so that $|V_{s}|\leq\ell$, $s\in V_{s}$, $t\in V_{t}$, and the total weight of edges between $V_{s}$ and $V_{t}$ in $G$ is at most $r$. SBM-$s$-$t$-Cut is NP-hard [11].

A clique in an undirected graph is a subset of pairwise adjacent vertices.

Clique
Input:	An undirected graph $G$ and an integer $\ell$.
Question:	Does $G$ contain a clique of $\ell$ vertices?

Clique is a well-known NP-hard problem [30]. Moreover, it is W[1]-hard with respect to $\kappa$ even when restricted to regular graphs [31, 32, 33].

For a digraph $G$ and a subset $A^{\prime}$ of arcs, $G-A^{\prime}$ denotes the digraph obtained from $G$ by removing all arcs in $A^{\prime}$. For two disjoint subsets $X,Y\subseteq V(G)$, let ${{\textsf{cut}}}_{G}(X,Y)$ be the set of all arcs from $X$ to $Y$, i.e., ${{\textsf{cut}}}_{G}(X,Y)=\{(v,u)\in A(G):v\in X,u\in Y\}$. The decision version of Directed Minimum $s$-$t$ Cut (DM-$s$-$t$-Cut) is defined as follows.

DM-$s$-$t$-Cut
Input:	A digraph $G$, a weight function $w:A(G)\rightarrow\mathbb{R}_{\geq 0}$, two distinct vertices $s$ and $t$ in $G$, an integer $r$.
Question:	Is there a bipartition $(V_{s},V_{t})$ of $V(G)$ such that $s\in V_{s}$, $t\in V_{t}$, and $w({{\textsf{cut}}}_{G}(V_{s},V_{t}))\leq r$?

Equivalently, DM-$s$-$t$-Cut determines if there is a subset $A^{\prime}\subseteq A(G)$ so that $w(A^{\prime})\leq r$ and there is no directed path from $s$ to $t$ in $G-A^{\prime}$. It is known that DM-$s$-$t$-Cut is polynomial-time solvable [16]).

4 Energy-Saving Partition of DAG

This section presents our results for ESP-DAG. We first pinpoint the complexity boundary of the problem concerning the number of machines.

Theorem 1.

ESP-DAG is NP-hard when $k\geq 3$.

Proof.

We prove Theorem 1 by a reduction from Multiway Cut to ESP-DAG. Let $(G,w,T,r)$ be an instance of Multiway Cut, where $T=\{t_{1},t_{2},\dots,t_{k}\}$. Without loss of generality, assume that $k\geq 2$. To construct an instance of ESP-DAG, we first arbitrarily orient $G$ into a DAG (e.g., we arbitrarily fix a linear order on $V(G)$ and orient edges forwardly). Let $\overrightarrow{G}$ denote this DAG. It holds that $V(G)=V(\overrightarrow{G})$. Then, we construct a graph $G^{\prime}$ obtained from $\overrightarrow{G}$ by subdividing all arcs: for each arc $(v,u)\in A(\overrightarrow{G})$, we introduce one new vertex $a(v,u)$, add the arcs $(v,a(v,u))$ and $(a(v,u),u)$, and remove the arc $(v,u)$. Clearly, each newly introduced vertex $a(v,u)$ has exactly one outneighbor $u$ and has exactly one inneighbor $v$. Moreover, $G^{\prime}$ remains as a DAG.

Now we construct two functions $p:V(G^{\prime})\times[k]\rightarrow\mathbb{R}_{\geq 0}$ and $q:V(G^{\prime})\rightarrow\mathbb{R}_{\geq 0}$ as follows. For each $x\in V(G^{\prime})$, we define

$$p(x,i)=\begin{cases}+\infty,&\text{if }x=t_{j},j\in[k]\setminus\{i\},\\ 0,&\text{otherwise}.\\ \end{cases}$$ (3)

For each $v\in V(\overrightarrow{G})$, we define $q(v)=+\infty$, and for each newly introduced vertex $a(v,u)\in V(G^{\prime})\setminus V(\overrightarrow{G})$ where $(v,u)\in A(\overrightarrow{G})$, we define $q(a(v,u))=w(\{v,u\})$.

The instance of ESP-DAG is $(G^{\prime},p,q,r)$. The reduction clearly runs in polynomial time. In the following, we show its correctness. For each $v\in V(G)$, let

$$B(v)=\{a(v,u):(v,u)\in A(\overrightarrow{G})\}.$$

For $V^{\prime}\subseteq V(G)$, let $B(V^{\prime})=\bigcup_{v\in V^{\prime}}B(v)$, and let $B[V^{\prime}]=B(V^{\prime})\cup V^{\prime}$.

$(\Rightarrow)$ Assume that the instance of Multiway Cut is a Yes-instance, i.e., there is a partition of $V(G)$ into $k$ sets $V_{1}$, $V_{2}$, $\dots$, $V_{k}$ such that $t_{i}\in V_{i}$ for all $i\in[k]$, and the total weight of edges crossing the partition is at most $r$. Note that for every two disjoint $V^{\prime},V^{\prime\prime}\subseteq V(G)$, $B(V^{\prime})$ and $B(V^{\prime\prime})$ are disjoint. It follows that $(B[V_{i}])_{i\in[k]}$ is a partition of $V(G^{\prime})$. Let $f$ be the assignment function corresponding to this partition, i.e., for every $i\in[k]$ and every $x\in B[V_{i}]$, it holds that $f(x)=i$. We claim that $f$ is a Yes-witness to the instance of ESP-DAG constructed above. First, by the definition of $(B[V_{i}])_{i\in[k]}$ and Equality (3), $\sum_{x\in V(G^{\prime})}p(x,f(x))=0$ holds. Moreover, by the definition of $(B[V_{i}])_{i\in[k]}$, we have that

$$\begin{split}{\textsf{Tran}_{G^{\prime}}(f,q)}&=\sum_{i\in[k]}\sum_{x\in f^{-1}(i)}\sum_{j\in[k]\setminus\{i\}}q(x)\cdot{\bf{1}}_{G^{\prime}}(x,f^{-1}(j))\\ &=\sum_{i\in[k]}\sum_{v\in V_{i}}\sum_{\begin{subarray}{c}a(v,u)\in B(v),\\ j\in[k]\setminus\{i\},u\in f^{-1}(j)\end{subarray}}q(a(v,u))\\ &=\sum_{i\in[k]}\sum_{v\in V_{i}}\sum_{\begin{subarray}{c}(v,u)\in A(\overrightarrow{G}),\\ j\in[k]\setminus\{i\},u\in V_{j}\end{subarray}}q(a(v,u))\\ &=\sum_{\begin{subarray}{c}i,j\in[k],i\neq j,\\ v\in V_{i},u\in V_{j},\\ \{v,u\}\in E(G)\end{subarray}}w(\{v,u\})\leq r.\\ \end{split}$$ (4)

To see that Equality (4) holds, observe that for every $v\in V_{i}$, where $i\in[k]$, it holds that $B(v)\subseteq B[V_{i}]$ and $\Gamma^{+}_{G^{\prime}}{(v)}=B(v)$. Consequently, only outputs of newly introduced vertices $a(v,u)\in V(G^{\prime})\setminus V(\overrightarrow{G})$ necessitate transfer. Moreover, the outdegree of each newly introduced vertex $a(v,u)$ is exactly one. This ensures the correctness of the transition from the first line to the third line in Equality (4). The transition from the penultimate line to the last line in Equality (4) follows from the construction of $\overrightarrow{G}$ from $G$, and the definition of the function $q$. Now, we can conclude that the instance of ESP-DAG is a Yes-instance.

$(\Leftarrow)$ Assume that the instance of ESP-DAG is a Yes-instance, i.e., there is an assignment function $f:V(G^{\prime})\rightarrow[k]$ such that

$$\left(\sum_{x\in V(G^{\prime})}p(x,f(x))\right)+{\textsf{Tran}_{G^{\prime}}(f,q)}\leq r.$$ (5)

By Equality (3), for every terminal $t_{i}\in T$, where $i\in[k]$, it holds that $f(t_{i})=i$. As $p(x,i)=0$ for all $x\in V(G^{\prime})\setminus T$ and $i\in[k]$, we have that $\sum_{x\in V(G^{\prime})}p(x,f(x))=0$. For each $i\in[k]$, let $V_{i}=\{v\in V(G):f(v)=i\}$. We show below that the total weight of edges in $G$ crossing the partition $(V_{i})_{i\in[k]}$ is exactly ${\textsf{Tran}_{G^{\prime}}(f,q)}$, which is at most $r$. To this end, recall that $q(v)=+\infty$ for every $v\in V(G)$. Then, by Inequality (5), we know that $f(v)=f(x)$ for all $x\in B(v)$ and $v\in V(G)$. That is, for every $v\in V(G)$, $v$ and all its outneighbors in $G^{\prime}$ are assigned the same value by $f$. To be more precise, for every $i\in[k]$, it holds that $B[V_{i}]\subseteq f^{-1}(i)$. As $B[V_{i}]$ and $B[V_{j}]$ are disjoint as long as $V_{i}$ and $V_{j}$ are, this indeed means that $B[V_{i}]=f^{-1}(i)$ for all $i\in[k]$. Then, by the same reasoning in the proof for the $(\Rightarrow)$ direction, we infer that Equality (4) holds too in this direction. Therefore, the instance of Multiway Cut is a Yes-instance. ∎

We remark that as the optimization version of Multiway Cut is APX-hard for every $k\geq 3$ [29], our reduction in the proof of Theorem 1 indicates that the optimization version of ESP-DAG is APX-hard for every $k\geq 3$. Moreover, as Multiway Cut remains NP-hard for every $k\geq 3$ even when all edges have the same weight [29], our reduction also implies that ESP-DAG remains NP-hard for every $k\geq 3$, even when the two functions $p$ and $q$ each has two different values, with one value being identical. Furthermore, astute readers may observe that our reduction can be readily adapted to demonstrate an even more compelling result:

Corollary 1.

ESP-DAG is NP-hard for every $k\geq 3$ even when the two functions $p$ and $q$ have overall two different values.

The proof of the result can be done by replacing the value $0$ in Equality (3) with $1$, resetting $q(a(v,u))=1$ for all newly introduced vertices $a(v,u)$, and resetting $r\coloneqq r+|V|$ in the instance of ESP-DAG constructed in our reduction.

When $p$ is a constant function, ESP-DAG can be solved trivially. Therefore, we have a complexity dichotomy for ESP-DAG with respect to the number of different values of the function $p$. Towards a complexity dichotomy for ESP-DAG concerning the number of machines, we have the following result.

Theorem 2.

ESP-DAG with $k=2$ is linear-time reducible to DM-$s$-$t$-Cut.

Proof.

Let $(G,p,q,r)$ be an instance of ESP-DAG where $G$ is a DAG, and $p:V(G)\times[2]\rightarrow\mathbb{R}_{\geq 0}$ and $q:V(G)\rightarrow\mathbb{R}_{\geq 0}$ are two functions. Our reduction is as follows. We first construct a digraph $G^{\prime}$ obtained from $G$ by performing the following operations:

1. For each nonsink $v\in V(G)$, we perform the following operations:

   1. create two vertices $v^{+}$ and $v^{-}$;

   2. for every outneighbor $u$ of $v$ in $G$, add the arcs ${(v^{+},u)}$ and $(u,v^{-})$, and let the weights of both arcs be $+\infty$;

   3. add the arcs $(v,v^{+})$ and $(v^{-},v)$, and let the weights of both arcs be $q(v)$;

   4. remove all arcs from $v$ to all its outneighbors except $(v,v^{+})$.

2. Create two vertices $s$ and $t$, add arcs from $s$ to all vertices in $V(G)$, and add arcs from all vertices in $V(G)$ to $t$. For each $v\in V(G)$, let the weight of the arc $(s,v)$ be $p(v,2)$, and let that of $(v,t)$ be $p(v,1)$. (Note that there are no arcs between the newly created vertices and $\{s,t\}$.)

We refer to Figure 1 for an illustration of the construction of $G^{\prime}$.

Figure 1: An illustration of the construction of the graph $G^{\prime}$ from a DAG $G$ in the proof of Theorem 2. Red arcs have positive infinite weights. For clarity, the weights of other arcs are not shown.

Let $w:A(G^{\prime})\rightarrow\mathbb{R}_{\geq 0}$ be the function such that for each arc $e\in A(G^{\prime})$ it holds that $w(e)$ equals the weight of $e$ defined above. The instance of DM-$s$-$t$-Cut is $(G^{\prime},w,s,t,r)$. The reduction clearly can be carried out in time $O(n+m)$, where $n$ and $m$ are respectively the number of vertices and the number of edges of $G$. It remains to prove its correctness.

$(\Rightarrow)$ Assume that there are two disjoint $V_{1},V_{2}\subseteq V(G)$ such that

$$\left(\sum_{v\in V_{i},i\in[2]}p(v,i)\right)+{\textsf{BTran}_{G}((V_{1},V_{2}),q)}\leq r.$$ (6)

Below we construct a partition $(X,Y)$ of $V(G^{\prime})$ such that $s\in X$, $t\in Y$, and the total weight of edges crossing $(X,Y)$ is at most $r$. To achieve this, we define the following sets:

• For each $i\in[2]$, let $V_{i}^{i}=\{v\in V(G):\Gamma^{+}_{G}{[v]}\subseteq V_{i}\}$ be the set of vertices $v\in V(G)$ such that $v$ and all its outneighbors in $G$ are in the same set $V_{i}$. (Notice that each isolated vertex of $G$ is either in $V_{1}^{1}$ or in $V_{2}^{2}$.)

• For each $i\in[2]$, let $V_{i}^{i\star}=\{v^{+}:v\in V_{i}^{i},\Gamma^{+}_{G}{(v)}\neq\emptyset\}\cup\{v^{-}:v\in V_{i}^{i},\Gamma^{+}_{G}{(v)}\neq\emptyset\}$ be the set of vertices constructed for vertices from $V_{i}^{i}$ by Operation (1a).

• For each $i,j\in[2]$ such that $i\neq j$, let $V_{i}^{j}=\{v\in V_{i}:\Gamma^{+}_{G}{(v)}\cap V_{j}\neq\emptyset\}$ be the set of vertices from $V_{i}$ having at least one outneighbor from $V_{j}$ in $G$.

• For each $i,j\in[2]$ such that $i\neq j$, let $V_{i}^{j+}=\{v^{+}:v\in V_{i}^{j}\}$.

• For each $i,j\in[2]$ such that $i\neq j$, let $V_{i}^{j-}=\{v^{-}:v\in V_{i}^{j}\}$.

The above defined ten sets are pairwise disjoint, and their union is exactly $V(G^{\prime})\setminus\{s,t\}$. Let

$$X=V_{1}^{1}\cup V_{1}^{1\star}\cup V_{1}^{2}\cup V_{1}^{2-}\cup V_{2}^{1-}\cup\{s\},$$

and let

$$Y=V_{2}^{2}\cup V_{2}^{2\star}\cup V_{2}^{1}\cup V_{1}^{2+}\cup V_{2}^{1+}\cup\{t\}.$$

It is not difficult to verify that $X\cap Y=\emptyset$ and $X\cup Y=V(G^{\prime})$. To conclude this part of the proof, we need to demonstrate that the total weight of edges crossing the bipartition $(X,Y)$ in $G^{\prime}$ is at most $r$. To achieve this, we identify the edges crossing this bipartition, relying on Claims 1 through 5 presented below.

Claim 1.

None of $V_{1}^{1}\cup V_{1}^{1\star}\cup V_{1}^{2-}$ has any outneighbors from $Y\setminus\{t\}$ in $G^{\prime}$.

Proof.

By the construction of $G^{\prime}$ and the definitions of $V_{1}^{1}$ and $V_{1}^{1\star}$, all outneighbors of vertices of $V_{1}^{1}$ in $G^{\prime}$ are contained in $V_{1}^{1\star}\cup\{t\}$. As $V_{1}^{1\star}\subseteq X$, none of $V_{1}^{1}$ has any outneighbors from $Y\setminus\{t\}$ in $G^{\prime}$.

For a vertex $v^{-}\in V_{1}^{1\star}$ where $v\in V_{1}^{1}$, the only outneighbor of $v^{-}$ in $G^{\prime}$ is $v$ which is clearly not in $Y$. For a vertex $v^{+}\in V_{1}^{1\star}$ where $v\in V_{1}^{1}$, the outneighbors of $v^{+}$ in $G^{\prime}$ are exactly the outneighbors of $v$ in the graph $G$ which are all contained in $V_{1}^{1}$. As $V_{1}^{1}\subseteq X$, none of the outneighbors of $v^{+}$ is from $Y$.

By the definition of $V_{1}^{2-}$, every $v^{-}\in V_{1}^{2-}$ has exactly one outneighbor $v$ in $G^{\prime}$ which is from $V_{1}^{2}$. Clearly, $V_{1}^{2}$ and $Y$ are disjoint. ∎

Claim 2.

All arcs from $V_{1}^{2}$ to $Y\setminus\{t\}$ in $G^{\prime}$ are contained in ${{\textsf{cut}}}_{G^{\prime}}(V_{1}^{2},V_{1}^{2+})$.

Proof.

By the construction of $G^{\prime}$, every $v\in V(G)$ has exactly two outneighbors $v^{+}$ and $t$ in $G^{\prime}$. By the definition of $V_{1}^{2}$, for every $v\in V_{1}^{2}$, $v^{+}$ is contained in $V_{1}^{2+}$. ∎

Claim 3.

None of $V_{2}^{2}\cup V_{2}^{2\star}\cup V_{2}^{1+}$ has any inneighbors from $X\setminus\{s\}$ in $G^{\prime}$.

Proof.

Let $v$ be any arbitrary vertex in $V_{2}^{2}$, if there are any. We consider first the case where $v$ is neither a sink nor a source of $G$. By the construction of $G^{\prime}$, it holds that $\Gamma^{-}_{G^{\prime}}{(v)}\setminus\{s\}=\{v^{-}\}\cup\{u^{+}:u\in\Gamma^{-}_{G}{(v)}\}$ (Operations (1c) and (1d)). Clearly, $v^{-}\in V_{2}^{2\star}$. Let $u$ be a vertex in $\Gamma^{-}_{G}{(v)}$. If $u\in X$, then as $u\in V_{1}^{2}$ and $v\in Y$, we know that $u\in V_{1}^{2}$. It follows that $u^{+}\in V_{1}^{2+}$. If $u\in Y$, then either $u^{+}\in V_{2}^{2\star}$ (when $u\in V_{2}^{2}$) or $u^{+}\in V_{2}^{1+}$ (when $u\in V_{2}^{1}$). In both cases, we have that $u^{+}\in Y$. Therefore, none of the inneighbors of $v$ in $G^{\prime}$ are from $X\setminus\{s\}$. In the case where $v$ is a sink but not a source of $G$, we have that $\Gamma^{-}_{G^{\prime}}{(v)}\setminus\{s\}=\{u^{+}:u\in\Gamma^{-}_{G}{(v)}\}$, in the case where $v$ is a source but not a sink of $G$, we have that $\Gamma^{-}_{G^{\prime}}{(v)}\setminus\{s\}=\{v^{-}\}$, and in the case where $v$ is an isolated vertex in $G$, we have that $\Gamma^{-}_{G^{\prime}}{(v)}\setminus\{s\}=\emptyset$. The proof for that none of the inneighbors of $v$ in $G^{\prime}$ is from $X\setminus\{s\}$ in these cases are the same as above.

Consider now a vertex $v^{+}\in V_{2}^{2\star}$ where $v\in V_{2}^{2}$. By the construction of $G^{\prime}$, we know that $v$ is the only inneighbor of $v^{+}$ in $G^{\prime}$. For a vertex $v^{-}\in V_{2}^{2\star}$ where $v\in V_{2}^{2}$, by the construction of $G^{\prime}$, we know that $\Gamma^{-}_{G^{\prime}}{(v^{-})}=\Gamma^{+}_{G}{(v)}$. As $v\in V_{2}^{2}$, we know that $\Gamma^{+}_{G}{(v)}\subseteq V_{2}$. Clearly, $V_{2}$ is the union of $V_{2}^{1}$ and $V_{2}^{2}$ which are both contained in $Y$. We can conclude now that none of the inneighbors of any vertex in $V_{2}^{2\star}$ is from $X\setminus\{s\}$.

Now we prove the claim for $V_{2}^{1+}$. Let $v^{+}$ be a vertex from $V_{2}^{1+}$. Obviously, $v\in V_{2}^{1}$. By the construction of $G^{\prime}$, $v$ is the only inneighbor of $v^{+}$. As $V_{2}^{1}\subseteq Y$, the claim holds. ∎

Claim 4.

All arcs from $V_{2}^{1-}$ to $Y$ in $G^{\prime}$ are contained in ${{\textsf{cut}}}_{G^{\prime}}(V_{2}^{1-},V_{2}^{1})$.

Proof.

By the construction of $G^{\prime}$, there are no arcs from $V_{2}^{1-}$ to $V_{1}^{2+}\cup\{t\}$ in $G^{\prime}$. By Claim 3, there are no arcs from $V_{2}^{1-}$ to $V_{2}^{2}\cup V_{2}^{2\star}\cup V_{2}^{1+}$ in $G^{\prime}$. This leaves only the possibility that all arcs from $V_{2}^{1-}$ to $Y$ in $G^{\prime}$ are contained in ${{\textsf{cut}}}_{G^{\prime}}(V_{2}^{1-},V_{2}^{1})$. ∎

The following claim is a consequence of Claims 1–4.

Claim 5.

${{\textsf{cut}}}_{G^{\prime}}(X,Y)$ is the union of the following pairwise disjoint sets:

• ${{\textsf{cut}}}_{G^{\prime}}(\{s\},Y)$, i.e., the set of all arcs from $s$ to all vertices of $Y$ in the graph $G^{\prime}$.

• ${{\textsf{cut}}}_{G^{\prime}}(X,\{t\})$, i.e., the set of all arcs from all vertices in $X$ to $t$ in the graph $G^{\prime}$.

• ${{\textsf{cut}}}_{G^{\prime}}(V_{1}^{2},V_{1}^{2+})$, i.e., the set of arcs from $V_{1}^{2}$ to $V_{1}^{2+}$ in the graph $G^{\prime}$.

• ${{\textsf{cut}}}_{G^{\prime}}(V_{2}^{1-},V_{2}^{1})$, i.e., the set of arcs from $V_{2}^{1-}$ to $V_{2}^{1}$ in the graph $G^{\prime}$.

Armed with Claim 5, we examine the size of the cut from $X$ to $Y$ in $G^{\prime}$. By Operation (2) in the construction of $G^{\prime}$, we know that there is an arc from $s$ to every vertex in $V_{2}^{1}\cup V_{2}^{2}$, and there is no arc from $s$ to any other vertices in $Y\setminus(V_{2}^{1}\cup V_{2}^{2})$. Then, by Operation (1c) in the construction of $G^{\prime}$, we know that

$$w({{\textsf{cut}}}_{G^{\prime}}(\{s\},Y))=\sum_{v\in V_{2}^{1}\cup V_{2}^{2}}w((s,v))=\sum_{v\in V_{2}^{1}\cup V_{2}^{2}}p(v,2).$$

Analogously, we know that $w({{\textsf{cut}}}_{G^{\prime}}(X,\{t\}))=\sum_{v\in V_{1}^{1}\cup V_{1}^{2}}p(v,1)$. From $V_{1}=V_{1}^{1}\cup V_{1}^{2}$ and $V_{2}=V_{2}^{1}\cup V_{2}^{2}$, we obtain

$$w({{\textsf{cut}}}_{G^{\prime}}(\{s\},Y))+w({{\textsf{cut}}}_{G^{\prime}}(X,\{t\}))=\sum_{v\in V_{i},i\in[2]}p(v,i).$$ (7)

Now we analyze $w({{\textsf{cut}}}_{G^{\prime}}(V_{1}^{2},V_{1}^{2+}))$. As for every vertex $v^{+}\in V(G^{\prime})$ where $v\in V(G)$, $v$ is the only inneighbor of $v^{+}$ in $G^{\prime}$, we have that

$${{\textsf{cut}}}_{G^{\prime}}(V_{1}^{2},V_{1}^{2+})=\{(v,v^{+}):v\in V_{1}^{2}\}.$$

It follows that

$$w({{\textsf{cut}}}_{G^{\prime}}(V_{1}^{2},V_{1}^{2+}))=\sum_{v\in V_{1}^{2}}w((v,v^{+}))=\sum_{v\in V_{1}^{2}}q(v).$$

Analogously, we can obtain that $w({{\textsf{cut}}}_{G^{\prime}}(V_{2}^{1-},V_{2}^{1}))=\sum_{v\in V_{2}^{1}}q(v)$.

By Claim 1, and the facts that $V_{1}=V_{1}^{1}\cup V_{1}^{2}$, $V_{2}\subseteq Y$, and every vertex from $V_{1}^{2}$ has at least one outneighbor from $V_{2}$ in the graph $G$, we have that $\sum_{v\in V_{1}}q(v)\cdot{\bf{1}}_{G}(v,V_{2})=\sum_{v\in V_{1}^{2}}q(v)$. Similarly, it holds that $\sum_{v\in V_{2}}q(v)\cdot{\bf{1}}_{G}(v,V_{1})=\sum_{v\in V_{2}^{1}}q(v)$. We arrive at

$$w({{\textsf{cut}}}_{G^{\prime}}(V_{1}^{2},V_{1}^{2+}))+w({{\textsf{cut}}}_{G^{\prime}}(V_{2}^{1-},V_{2}^{1}))={\textsf{BTran}_{G}((V_{1},V_{2}),q)}.$$ (8)

From Claim 5, and Equalities (7) and (8), we have that

$$w({{\textsf{cut}}}_{G^{\prime}}(X,Y))=\left(\sum_{v\in V_{i},i\in[2]}p(v,i)\right)+{\textsf{BTran}_{G}((V_{1},V_{2}),q)},$$

which is at most $r$ by Inequality (6). It follows that the instance of DM-$s$-$t$-Cut is a Yes-instance.

$(\Leftarrow)$ Assume that there is a bipartition $(X,Y)$ of $V(G^{\prime})$ such that $s\in X$, $t\in Y$, and $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))\leq r$. Without loss of generality, we assume that $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$ is minimized among all bipartitions of $V(G^{\prime})$. We show that ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$ possesses several structural properties which help us construct a desired assignment of $V(G)$.

The first property, as formally articulated in the ensuing claim, affirms that when a vertex $v$ in $G$ is included in the same part of $(X,Y)$ alongside all its outneighbors in $G$, retaining both newly introduced vertices for $v$ within the same part as $v$ emerges as an optimal strategy.

Claim 6.

Let $v\in V$ such that $\Gamma^{+}_{G}{(v)}\neq\emptyset$ and all outneighbors of $v$ in $G$ are in the same $Z\in\{X,Y\}$. Let $\overline{Z}=V(G^{\prime})\setminus Z$. Let $X^{\prime}=Z\cup\{v^{+},v^{-}\}$ and let $Y^{\prime}=\overline{Z}\setminus\{v^{+},v^{-}\}$. Then, it holds that $w({{\textsf{cut}}}_{G^{\prime}}(X^{\prime},Y^{\prime}))\leq w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$ if $s\in X^{\prime}$, and $w({{\textsf{cut}}}_{G^{\prime}}(Y^{\prime},X^{\prime}))\leq w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$ if $s\in Y^{\prime}$.

Proof.

Let $v$ be as stipulated in Claim 6. We consider first the case where $\Gamma^{+}_{G}{[v]}\subseteq X$. Thus, $X^{\prime}=X\cup\{v^{+},v^{-}\}$ and $s\in X^{\prime}$ in this case. We need to prove that $w({{\textsf{cut}}}_{G^{\prime}}(X^{\prime},Y^{\prime}))\leq w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$. To this end, observe that none of $\{(u,v^{-}):u\in\Gamma^{+}_{G}{(v)}\}$ is in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$, due to the infinite weight of arcs in this set ((1b)). It follows that $v^{-}\in X$. If $v^{+}\in X$, it holds that $(X^{\prime},Y^{\prime})=(X,Y)$, and thus $w({{\textsf{cut}}}_{G^{\prime}}(X^{\prime},Y^{\prime}))=w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$; we are done. Assume now that $v^{+}\in Y$. Then, as $\Gamma^{-}_{G^{\prime}}{(v^{+})}=\{v\}$, we have that $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))-w({{\textsf{cut}}}_{G^{\prime}}(X^{\prime},Y^{\prime}))=w((v,v^{+}))\geq 0$.

We consider now the case where $\Gamma^{+}_{G}{[v]}\subseteq Y$. In this case, $X^{\prime}=Y\cup\{v^{+},v^{-}\}$, $Y^{\prime}=X\setminus\{v^{+},v^{-}\}$, and $s\in Y^{\prime}$. Observe that none of $\{(v^{+},u):u\in\Gamma^{+}_{G}{(v)}\}$ is in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$, due to the infinite weight of arcs in this set. It follows that $v^{+}\in Y$. If $v^{-}\in Y$, it holds that $(Y^{\prime},X^{\prime})=(X,Y)$, and thus $w({{\textsf{cut}}}_{G^{\prime}}(Y^{\prime},X^{\prime}))=w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$; we are done. Assume now that $v^{-}\in X$. Then, as $\Gamma^{+}_{G^{\prime}}{(v^{-})}=\{v\}$, we have that $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))-w({{\textsf{cut}}}_{G^{\prime}}(Y^{\prime},X^{\prime}))=w((v^{-},v))\geq 0$. ∎

By Claim 6, we may assume that, for every $v\in V$ such that $\Gamma^{+}_{G}{(v)}\neq\emptyset$, if $\Gamma^{+}_{G}{[v]}$ are contained in the same $Z\in\{X,Y\}$, then $\{v^{+},v^{-}\}$ are also contained in $Z$.

The subsequent property essentially posits that for any vertex $v$ of $G$ placed in part $X$, if at least one of $v$’s outneighbors in $G$ is placed in the opposite part $Y$, then the arc $(v,v^{+})$, whose weight equals the amount of energy consumption of transferring $v$, must necessarily be in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$.

Claim 7.

Let $v\in X\cap V(G)$ such that $\Gamma^{+}_{G}{(v)}\cap Y\neq\emptyset$. Then, $(v,v^{+})\in{{\textsf{cut}}}_{G^{\prime}}(X,Y)$ and none of the arcs entering or leaving $v^{-}$ is contained in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$.

Proof.

Let $v$ be as stipulated in the claim. Let $u$ be any arbitrary vertex from $\Gamma^{+}_{G}{(v)}\cap Y$. It must hold that $v^{+}\in Y$, since otherwise the arc $(v^{+},u)$ which has infinite weight is contained in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$, contradicting with $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))\leq r$. As $v$ is the only outneighbor of $v^{-}$ in $G^{\prime}$ and $v\in X$, the only arc $(v^{-},v)$ leaving $v^{-}$ is excluded from ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$. Finally, recall that $\Gamma^{-}_{G^{\prime}}{(v^{-})}=\{u\in V(G):u\in\Gamma^{+}_{G}{(v)}\}$, and by the definition of the function $w$, every arc $(u,v^{-})$ where $u\in\Gamma^{+}_{G}{(v)}$ has an infinite weight. As $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))\leq r$, we know that none of the arcs entering $v^{-}$ can be in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$. ∎

By Claim 7, for every $v\in X\cap V(G)$ which has at least one outneighbor in $G$ that is put in from $Y$, we may assume that $v^{+}\in Y$. Moreover, as $v$ is the only outneighbor of $v^{-}$ in $G^{\prime}$, we may assume that $v^{-}\in X$.

Applying analogous reasoning, we derive the following assertion.

Claim 8.

Let $v\in Y\cap V(G)$ such that $\Gamma^{+}_{G}{(v)}\cap X\neq\emptyset$. Then, $(v^{-},v)\in{{\textsf{cut}}}_{G^{\prime}}(X,Y)$, and none of the arcs entering or leaving $v^{+}$ is contained in ${{\textsf{cut}}}_{G^{\prime}}(X,Y)$.

For every $v\in Y\cap V(G)$ as in Claim 8, we may assume that $v^{-}\in X$ and $v^{+}\in Y$.

Let $V_{1}=V(G)\cap X$ and let $V_{2}=V(G)\cap Y$. We show below that the assignment corresponding to $(V_{1},V_{2})$ is a Yes-witness of the instance of ESP-DAG. By the definition of the function $w$, we have that

$$\sum_{v\in V_{1}}p(v,1)+\sum_{v\in V_{2}}p(v,2)=\sum_{v\in V_{1}}w((v,t))+\sum_{v\in V_{2}}w((s,v)).$$ (9)

Let $V_{1}^{\prime}=\{v\in V_{1}:\Gamma^{+}_{G}{(v)}\cap Y\neq\emptyset\}$ and let $V_{2}=\{v\in V_{2}:\Gamma^{+}_{G}{(v)}\cap X\neq\emptyset\}$. By Claims 6–8, we have:

• For every $i\in[2]$ and every $v\in V_{i}\setminus V_{i}^{\prime}$, $\{v^{+},v^{-}\}$ are contained in the same part of $(X,Y)$ as $v$.

• For every $v\in V_{1}^{\prime}\cup V_{2}^{\prime}$, $v^{+}\in Y$ and $v^{-}\in X$.

The following equalities follow:

$$\sum_{v\in V_{1}}q(v)\cdot{\bf{1}}_{G}(v,V_{2})=\sum_{v\in V_{1}^{\prime}}w((v,v^{+})),$$ (10)

$$\sum_{v\in V_{2}}q(v)\cdot{\bf{1}}_{G}(v,V_{1})=\sum_{v\in V_{2}^{\prime}}w((v^{-},v)).$$ (11)

The sum of the left sides of Equalities (9)–(11) is $\left(\sum_{v\in V_{i},i\in[2]}p(v,i)\right)+{\textsf{BTran}_{G}((V_{1},V_{2}),q)}$, and the sum of the right sides of them is $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))$. From $w({{\textsf{cut}}}_{G^{\prime}}(X,Y))\leq r$, we know that the instance of ESP-DAG is a Yes-instance. ∎

As DM-$s$-$t$-Cut can be solved in time $O(n\cdot m\cdot\log(n^{2}/m))$ [34], by Theorem 2, we reach the following result.

Corollary 2.

ESP-DAG with $k=2$ can be solved in time $O(n\cdot m\cdot\log(n^{2}/m))$, where $n$ and $m$ are respectively the number of vertices and the number of edges of the input graph.

Next, we derive a pseudo polynomial-time algorithm for a special case of ESP-DAG where the input DAG is a directed path. Tasks having such precedence dependencies are relevant to many applications (see, e.g., [35, 36]).

Theorem 3.

ESP-DAG can be solved in time $O(n\cdot k^{3}\cdot M)$ when the given DAG is a directed path of $n$ vertices, where $M$ is the number of bits to encode any number in the ranges of the functions $p$ and $q$.

Proof.

Let $I=(G,p,q,r)$ be an instance of ESP-DAG, where $G$ is a directed path, $p:V(G)\times[k]\rightarrow\mathbb{R}_{\geq 0}$ and $q:V(G)\rightarrow\mathbb{R}_{\geq 0}$ are two functions, and $r$ is a number. Let $M$ be the maximum value of the functions $p$ and $q$. We derive a dynamic programming algorithm to solve the problem as follows. Let $(v_{1},v_{2},\dots,v_{n})$ be the directed path representing $G$. We maintain a table $S(i,j)$ where $i\in[n]$ and $j\in[k]$. Specifically, $S(i,j)$ is defined as the value of an optimal solution to the instance $I$ restricted to the first $i$ vertices under the restriction that $v_{i}$ is assigned to the machine $j$:

$$S(i,j)=p(v_{i},j)+\min_{\begin{subarray}{c}f:\{v_{1},\dots,v_{i}\}\rightarrow[k],\\ f(v_{i})=j\end{subarray}}\left\{\sum_{x\in[i-1]}p(v_{x},f(v_{x}))+\sum_{\begin{subarray}{c}x\in[i-1]~{}{\text{such that}}\\ f(v_{x})\neq f(v_{x+1})\end{subarray}}q(v_{x})\right\}.$$

By the definition, for each $j\in[k]$, we have that $S(1,j)=p(v_{1},j)$. We use the following recursion to update the table:

$$S(i,j)=p(v_{i},j)+\min\left\{S(i-1,j),\min_{j^{\prime}\in[k]\setminus\{j\}}S(i-1,j^{\prime})+q(v_{i-1})\right\}.$$

After the table is computed, we conclude that $I$ is a Yes-instance if and only if

$$\min_{j\in[k]}S(n,j)\leq r.$$

As there are at most $n\cdot k$ entries in the table, and computing each entry needs to check at most $k$ other entries and perform addition/comparison operations $O(k)$ times on numbers of maximum values $M$, the table can be computed in time $O(n\cdot k^{3}\cdot M)$. ∎

5 Sized-Bounded Energy-Saving Partition of DAG with Two Machines

Now we switch our focus to the natural variant SB-ESBP-DAG where there are only two machines and one of them is capable of executing only a limited number $\ell$ of tasks.

Recall that SBM-$s$-$t$-Cut is exactly Multiway Cut where $k=2$ with an additional restriction that one of the parts in the desired bipartition contains at most $\ell$ vertices. Then, based on the reduction in the proof of Theorem 1 and the fact that SBM-$s$-$t$-Cut is NP-hard, we know that SB-ESBP-DAG is NP-hard as well. his observation prompts an exploration into whether the natural parameter $\ell$ might offer a potential tractability insight within the realm of parameterized complexity theory.

A concise introduction to parameterized complexity theory can be beneficial for readers unfamiliar with the concept. A parameterized problem is a subset of $\Sigma^{*}\times\mathbb{N}$, where $\Sigma$ is a finite alphabet. A parameterized problem can be either fixed-parameter tractable (FPT) or fixed-parameter intractable. In particular, a parameterized problem is FPT if there is an algorithm which correctly determines for each instance $(I,\kappa)$ of the problem whether $(I,\kappa)$ is a Yes-instance in time $O(f(\kappa)\cdot|I|^{O(1)})$, where $f$ is a computable function and $|I|$ is the size of $I$. W[1]-hard problems are considered intractable, in the sense that unless FPT$=$W[1] (which is widely believed to be unlikely), they do not admit any FPT-algorithms. For greater details on parameterized complexity theory, we refer to [8, 9, 10].

Answering the above question, we show that SB-ESBP-DAG is W[1]-hard with respect to $\ell$. To this end, through a trivial reduction from Clique on regular graphs, we show that SBM-$s$-$t$-Cut is W[1]-hard with respect to $\ell$ even in a special case. Despite the widespread attention this problem has received, it is remarkable that such a W[1]-hardness result has not been previously documented. Therefore, our reduction bolsters the NP-hardness of SBM-$s$-$t$-Cut as examined in [11].

Theorem 4.

SBM-$s$-$t$-Cut is W[1]-hard with respect to $\ell$. Moreover, this holds even if the edges in the given graph have at most two different weights.

Proof.

Let $(G,\ell)$ be an instance of Clique where $G$ is a regular graph of $n$ vertices. Let $d$ denote the degree of vertices in $G$. We assume that $d>\ell$ since otherwise the problem is trivial to solve. Let $G^{\prime}$ be the graph obtained from $G$ by adding two vertices $s$ and $t$ both of which are adjacent to all vertices of $G$. Let $w:E(G^{\prime})\rightarrow\mathbb{R}_{\geq 0}$ be a function such that for every edge $\{x,y\}\in E(G^{\prime})$ it holds that

$$w(\{x,y\})=\begin{cases}d+2,&\text{if }s\in\{x,y\},\\ 1,&\text{otherwise}.\end{cases}$$

It is easy to verify that $G$ contains a clique of $\ell$ vertices if and only if there is a bipartition of $V(G^{\prime})$ such that $s$ and $t$ are in different parts with the one containing $s$ having $\ell+1$ vertices and, moreover, the total weight of edges crossing the bipartition is at most $(n-\ell)\cdot(d+2)+\ell\cdot(d-\ell+1)+\ell=n\cdot(d+2)-\ell^{2}$. ∎

By employing the reduction outlined in the proof of Theorem 1, Theorem 4 yields the following corollary.

Corollary 3.

SB-ESBP-DAG is W[1]-hard with respect to $\ell$. This holds even if the function $p$ has two different values and the function $q$ has three different values.

6 Conclusion

We have studied the complexity of ESP-DAG and its sized-bounded variant SB-ESBP-DAG. Our study offers several dichotomy results on the complexity of ESP-DAG, concerning the number $k$ of machines and the values of the function $p$. However, an intriguing open question persists: can SB-ESBP-DAG be solved in polynomial time when the function $q$ remains constant? Furthermore, the unresolved inquiry into whether SB-ESBP-DAG retains its status as NP-hard when $q$ exhibits only two distinct values adds another layer of complexity to this fascinating problem.