A Pseudopolynomial Algorithm to Minimize Maximum Lateness on Multiple Related Machines

Firstly, we note that on any given machine, any feasible ordering of the jobs done on that machine can also be done feasibly by doing all the jobs in order of increasing deadlines (as if $2$ consecutive jobs are out of order, we can swap them).[4] Also note that we can just scale the deadlines and the time it takes to do each job by the same amount and not change the problem, so we can assume the numbers to be integers.

Now, given two machines, we can compute feasibility as follows:

We DP on time and number of jobs seen: for a time $t$ and being given the first $i$ jobs, we want to have $D_{t,i}=1$ iff there is a feasible scheduling that makes one of the machines finish at exactly time $t$ while processing all of the first $i$ jobs on time and $D_{t,i}=0$ otherwise.

To determine the value of $D_{t,i}$, there must either be a scheduling where we put the $i^{th}$ job at the end of the schedule of the machine that ends up completing at time $t$ or in the other machine. To do this, we check $D_{t-t_{i},i-1}$ and $D_{t,i-1}$. If $D_{t-t_{i},i-1}=1$ and $t\leq d_{i}$, then we can add the $i^{th}$ job to the end of that machine and get a feasible scheduling of the first $i$ machines. If not, adding $i$ to the end of that machine does not give us a feasible scheduling. Otherwise, if $D_{t,i-1}=1$, then we know that the completion time of the other machine on the first $i-1$ jobs is

and that the other machine completes all of these jobs feasibly. If this value is at most $d_{i}-t_{i}$, then we can add the $i^{th}$ job to the end of it to get a valid scheduling.

Thus, we let $D_{t,i}=1$ if either of the above conditions are true and $0$ otherwise. This is correct since there are only two machines, and one of them has to complete at time $t$ for this to be $1$. The above checks both possibilities for whether they are possible, and so if neither works, then we know this is impossible.

Now, to check feasibility, we just need to look at the $n^{th}$ row and see if any of the values are $1$. If any are, we return $1$. Otherwise, we return $0$.

The runtime of this algorithm is as follows:

We know that the columns of the DP table only need to go up to $\sum_{i=1}^{n}t_{n}\leq nT$ where $T=\max t_{n}$ so we can make our table $n\times nT$.

Now, to fill out the $i^{th}$ row, we first compute $\left(\sum_{j=1}^{i-1}t_{j}\right)$ in $O(n)$ time. There are $n$ rows so all of these together takes $O(n^{2})$ time. Then, to fill in the value at each cell in that row, we check two values in the row above and then sometimes compute whether $\left(\sum_{j=1}^{i-1}t_{j}\right)-t\leq d_{i}-t_{i}$. Since we know the value of $\left(\sum_{j=1}^{i-1}t_{j}\right)$, each of these takes $O(1)$ time so we spend $O(1)$ time on each cell. There are $n^{2}T$ cells so this takes $O(n^{2}T)$ time.

Thus, the total runtime to check feasibility is $O(n^{2}T)$ and the runtime to solve $P2\mid L_{\max}$ is $O(n^{2}T\log(nT))$

To generalize to $m$ machines, we number the machines $1,2,\dots m$ and let $DP[i][(x_{1},x_{2},\cdots,x_{m-1})]$ indicate whether there is a feasible scheduling of the first $i$ jobs jobs such that machine $j$ completes all jobs scheduled to it in exactly $x_{j}$ time for $1\leq j\leq m-1$ (and this uniquely determines the completion time for machine $m$ since it’s just the remaining time that isn’t accounted for in the first $m-1$ machines).

We again assume that $t_{i}$ is the amount of time it takes to complete job $i$ and $d_{i}$ is the deadline of job $i$. To compute $DP[i][(x_{1},x_{2},\cdots,x_{m-1})]$, if $DP[i-1][(x_{1},x_{2},\cdots,x_{j}-t_{i},\cdots,x_{m-1}]=1$ for any $j$ such that $x_{j}\leq d_{i}$, then we have a feasible scheduling of the first $i-1$ jobs such that if we add the $i^{th}$ job to the end of the schedule for the $j^{th}$ machine, we get the desired completion times for all of the machines and it is still feasible.

In the other case, we add the $i^{th}$ job to the last machine. Then, we need to check that $DP[i-1][(x_{1},x_{2},\cdots,x_{m-1})]=1$, so there is a feasible scheduling of the first $i-1$ jobs with the same weights on the first $m-1$ machines. Then, for this scheduling we also need to make sure that adding the $i^{th}$ job to the last machine does not go over the deadline for that job. The last machine runs in time $\sum_{k=1}^{i}t_{k}-\sum_{k=1}^{m-1}x_{k}$, and we can make this transition as long as $\sum_{k=1}^{i}t_{k}-\sum_{k=1}^{m}x_{k}\leq d_{i}$.

Again, we can compute and store $\sum_{k=1}^{i}t_{k}$ each time we get to a new row in $O(n^{2})$ time. Then, each dimension is at most $nT$ where $T$ is the maximum time it takes to complete a job, so there are $n^{m}T^{m-1}$ total states, and computing the value of a state takes $O(m)$ time (as we need to check a transition for every single machine and potentially add up $m$ numbers), so the overall runtime is $O(mn^{m}T^{m-1})$ time, which is pseudopolynomial if $m$ is a constant.

This time, we will suppose that we are given for machine $i$ a rate parameter $\lambda_{i}$ which is the number of seconds it takes the machine to do $1$ unit of work. Just like above, we will suppose that job $i$ takes $t_{i}$ work and has a deadline of $d_{i}$. We will first scale the rate parameters and the amount of work each job takes until the work each job takes is an integer for every job since scaling both by the same amount doesn’t change anything. Then, for the rate parameters and deadlines, if we scale both of these by the same amount then we don’t change anything about the problem so we if we assume these numbers are all integers, our algorithm is pseudopolynomial in both the job lengths and machine rates.

Again, we let $DP[i][(x_{1},x_{2},\cdots,x_{m-1})]$ indicate whether there is a feasible arrangement of jobs such that machine $i$ runs in exactly $x_{i}$ time but here compute the value of $DP[i][(x_{1},x_{2},\cdots,x_{m-1})]$ as follows:

if $DP[i-1][(x_{1},x_{2},\cdots,x_{j}-t_{i}\lambda_{j},\cdots,x_{m-1}]=1$ such that $x_{j}\leq d_{i}$ for any $1\leq j\leq m-1$, we can add the $i^{th}$ job to the end of the schedule of the $j^{th}$ machine and get a feasible scheduling just like above. Otherwise, if we add it to the $m^{th}$ machine, we know that the machine runs in time

as from how we defined rates, if a machine has rate $\lambda_{k}$ and runs for $x_{k}$ time, then the total amount of work it does is $\frac{x_{k}}{\lambda_{k}}$. The total amount of work to do given the first $i$ jobs is $\sum_{k=1}^{i}t_{k}$, so we can subtract these values to get the amount of work that the last machine does and multiply by its rate to find the time it takes to complete all these jobs. Then, we just need to check if this value is less than or equal to $d_{i}$ to check if this is a valid scheduling since we know that the scheduling of the first $i-1$ jobs is feasible by induction. As there are the same number of states as in the $Pm\mid L_{\max}$ case and each state still takes $O(m)$ time to compute, the runtime is still $O(mn^{m}T^{m-1})$ and thus pseudopolynomial.

Since Bin Packing is strongly NP-hard, this means that $P\mid L_{\max}$ is also strongly NP-hard so if we do not assume there are a constant number of machines, we cannot even get a pseudopolynomial algorithm.