Online Generalized Scheduling

• Online generalized scheduling is a family of optimization problems where jobs, vectors, or requests arrive over time and must be assigned without full future information.
• The framework extends classical makespan minimization to include vector loads, polyhedral constraints, submodular objectives, and group-based completion measures.
• Analyses leverage competitive analysis, speed augmentation, and learning-augmented strategies to achieve provable performance guarantees under diverse operational constraints.

Searching arXiv for recent and foundational papers on online generalized scheduling and closely related models. Online generalized scheduling is a family of online optimization problems in which jobs, vectors, intervals, requests, or task blocks arrive over time and must be scheduled or assigned without full knowledge of future input. Relative to classical online makespan minimization on identical machines, the generalizations represented in the research literature include vector-valued loads over multiple resources, unrelated or heterogeneous machines, polyhedral and polymatroid feasibility regions, group-based completion objectives, nondecreasing job- and machine-dependent cost functions, replenishment and busy-time costs, cardinality constraints, non-clairvoyance, and prediction- or learning-augmented decision rules (Im et al., 2014, Lindermayr et al., 29 Jan 2025, Hathcock et al., 2024, Etesami, 2019). The unifying analytic theme is competitive analysis, often supplemented by speed augmentation, migration, regret, prediction error, or queue-stability guarantees.

1. Classical baseline and principal generalization axes

The classical baseline is online scheduling for makespan minimization: jobs arrive one by one and must be assigned immediately and irrevocably to machines, with objective $C_{\max}$. For identical machines, Graham’s List Scheduling assigns each job to the currently least loaded machine and has competitive ratio $2-\frac{1}{m}$; more refined deterministic and randomized bounds are known, but the exact competitive ratio for general $m$ remains unresolved within the ranges summarized in the literature (Dwibedy et al., 2020). At the same time, the classical problem admits a competitive-ratio approximation scheme: for $P||C_{\max}$, one can compute in polynomial time an online algorithm whose competitive ratio is at most $(1+\varepsilon)$ times the optimal online competitive ratio, using a finite adversary–scheduler game over rounded “trimmed-scenarios” (Chen et al., 2013).

What makes the problem “generalized” is the extension of at least one of the standard dimensions of the model. The literature represented here expands the input space from scalar jobs to vectors, blocks of dependent tasks, intervals, and graph or matroid requests; expands the machine model from identical to related, unrelated, or polyhedral resource systems; and expands the objective from makespan to $L_r$ load norms, generalized completion time, generalized flow time, busy time, replenishment-plus-flow-time, or weighted group completion (Im et al., 2014, Szalkai et al., 2015, Lindermayr et al., 29 Jan 2025, Györgyi et al., 2022). Some models further relax or refine commitment structure: online Minimum Peak Appointment Scheduling fixes appointment times online but defers room assignment to an offline phase, while prediction-augmented interval scheduling and streaming systems incorporate imperfect lookahead (Smedira et al., 2021, Boyar et al., 2023, Huang et al., 2020).

A compact way to view the landscape is through the following representative formulations.

Formulation	Defining feature	Representative result
Online vector scheduling	Jobs have $d$-dimensional loads	Tight ratios $\Theta(\log d/\log\log d)$ on identical machines and $\Theta(\log m+\log d)$ on unrelated machines (Im et al., 2014)
Polytope scheduling with groups	Objective $\min \sum_{S} w_S C_S$, $2-\frac{1}{m}$0	$2-\frac{1}{m}$1-competitive non-clairvoyant algorithm, tight (Lindermayr et al., 29 Jan 2025)
Online SAP	Monotone submodular feasibility constraints	Fractional $2-\frac{1}{m}$2-competitive algorithm (Hathcock et al., 2024)
General cost scheduling	Objective $2-\frac{1}{m}$3	Speed-augmented algorithms via optimal control and dual fitting (Etesami, 2019)

2. Multidimensional load balancing: vector scheduling and generalized load balancing

Vector scheduling is a central generalized model in which each job has a vector load rather than a scalar processing requirement. It was introduced by Chekuri and Khanna as a generalization of classical load balancing, motivated by modern data centers in which requests simultaneously consume processor cycles, storage space, and network bandwidth (Im et al., 2014). The 2014 resolution of the online complexity established that, for identical machines, the optimal competitive ratio is

$2-\frac{1}{m}$4

and for unrelated machines it is

$2-\frac{1}{m}$5

These bounds extend from makespan, viewed as an $2-\frac{1}{m}$6 objective, to general $2-\frac{1}{m}$7 norms. The paper explicitly characterizes the vector setting as exhibiting a strict logarithmic separation from scalar load balancing, and for identical machines it shows that the gap between online and offline for makespan is $2-\frac{1}{m}$8 (Im et al., 2014).

The lower-bound machinery is itself a substantial part of the theory. For identical machines, the tight lower bound is obtained through an online lower bound for the minimum monochromatic clique problem, via a novel online coloring game combined with a randomized coding scheme. The upper-bound side uses online assignment rules adapted to vector loads and $2-\frac{1}{m}$9 objectives. For unrelated machines and general $m$0 norms, the analysis requires a carefully constructed potential function that balances individual $m$1 objectives with an overall “convexified min-max” objective; the algorithm assigns each arriving job so as to minimize or control the increase in this potential (Im et al., 2014).

A precursor to the tight theory is the reduction from online vector scheduling to generalized load balancing. That reduction yields the first non-trivial online algorithm for vectors arriving online: each vector is assigned to minimize the $m$2 norm of the resulting load, giving an $m$3-approximation, and showing that generalized load balancing has no constant approximation algorithms running in polynomial time unless $m$4 (Zhu et al., 2012). In historical terms, this reduction-based result provided a logarithmic online guarantee, while the later tight bounds resolved the exact online complexity up to constant factors (Zhu et al., 2012, Im et al., 2014).

3. Polymatroid, submodular, and group-completion formulations

A broader generalization replaces per-machine scalar or vector capacity constraints by combinatorial feasibility structures. The Online Submodular Assignment Problem introduces a monotone submodular function $m$5 as the budget constraint over the ground set $m$6, with arriving parts $m$7 and assignment restrictions $m$8. The objective is

$m$9

subject to

$P||C_{\max}$0

This strictly generalizes GAP and captures laminar AdWords, online matroid coloring, arboricity, flow scheduling, and coflows. The paper gives a deterministic fractional $P||C_{\max}$1-competitive algorithm, which is provably optimal even in special cases, along with integral results under small bids and for matroid-rank utilities (Hathcock et al., 2024).

The central structural tool in SAP is the “water level” vector for polymatroids. It generalizes the classical water-filling paradigm and admits combinatorial, saddle-point, and market-equilibrium characterizations. The properties highlighted in the analysis are monotonicity and continuity, a feasibility test $P||C_{\max}$2 for all $P||C_{\max}$3, locality under small changes, and a duality relation to the Lovász extension,

$P||C_{\max}$4

These properties support incremental online pricing and the comparison of primal and dual progress (Hathcock et al., 2024).

A different abstraction generalizes the objective rather than the constraint system. In polytope scheduling with groups, jobs have processing requirements $P||C_{\max}$5, the feasible instantaneous rates satisfy

$P||C_{\max}$6

and the objective is

$P||C_{\max}$7

This simultaneously generalizes makespan and sum of weighted completion times: a single group $P||C_{\max}$8 recovers makespan, while singleton groups recover the classical min-sum objective (Lindermayr et al., 29 Jan 2025). In the non-clairvoyant online setting, the key algorithm uses Proportional Fairness with virtual weights

$P||C_{\max}$9

and at each time solves

$(1+\varepsilon)$0

The resulting competitive ratio is $(1+\varepsilon)$1, where $(1+\varepsilon)$2 is the size of the largest group, and a matching lower bound shows that $(1+\varepsilon)$3 is optimal (Lindermayr et al., 29 Jan 2025).

Taken together, these papers show that “generalized scheduling” often means one of two mathematically distinct moves: replacing machine feasibility by polymatroid or polyhedral structure, or replacing per-job objectives by set- or group-based objectives. A plausible implication is that many later models can be interpreted as combinations of these two moves.

4. General cost functions, primal–dual methods, and residual-optimum viewpoints

A major branch of the area studies online scheduling with objectives more general than makespan or linear flow time. On a single processor with preemption, generalized flow-time problems take the form

$(1+\varepsilon)$4

where $(1+\varepsilon)$5 and $(1+\varepsilon)$6 is non-decreasing. A primal–dual and dual-fitting framework interprets Highest-Density-First (HDF) as a primal–dual algorithm, proves HDF optimal for total fractional weighted flow time, and shows that the same analysis extends to generalized flow-time objectives, weighted completion time with general cost functions, and scheduling under polyhedral constraints (Angelopoulos et al., 2015). A notable methodological point is that the framework avoids potential functions and bypasses “black-box” rounding of fractional solutions, yielding improved competitive ratios in several settings.

For the broader objective

$(1+\varepsilon)$7

with job- and machine-dependent nonnegative nondecreasing cost functions $(1+\varepsilon)$8, an optimal-control framework provides a systematic way to construct dual variables. The model allows single-machine and unrelated-machine environments, preemption but no migration, and dispatch decisions upon arrival. For a single machine with special costs $(1+\varepsilon)$9, highest-density-first is optimal for the generalized fractional completion-time problem. For general nondecreasing costs, the paper gives a speed-augmented competitive algorithm; for multiple unrelated machines with convex $L_r$0, it gives a dispatch-and-reschedule algorithm whose guarantees depend on curvature parameters

$L_r$1

The stated bounds are $L_r$2-speed and $L_r$3-competitive on a single machine, and $L_r$4-speed and $L_r$5-competitive on unrelated machines (Etesami, 2019).

A more recent meta-algorithmic perspective replaces hand-designed rules by steepest descent on the residual offline optimum. Given remaining job sizes $L_r$6, the residual optimum is

$L_r$7

and the online rule chooses a feasible processing vector $L_r$8 minimizing $L_r$9. The resulting framework generalizes SRPT and yields scalable algorithms for minimizing weighted flow time whenever the residual optimum is supermodular. The paper states that this gives scalable results for matroid scheduling, generalized network flow, and generalized arbitrary speed-up curves, and that it provides the first non-trivial or scalable algorithm for many such problems (Chen et al., 2024).

These three viewpoints—dual dominance, optimal-control dual identification, and gradient descent on residual optimum—are technically different, but they converge on a common principle: generalized online scheduling is often tractable when the offline residual problem has enough convex, supermodular, or dual-structured regularity to support a continuously updated certificate.

5. Constraint-enriched and operationally coupled variants

Several generalized models add operational constraints that are absent from standard load-balancing or flow-time formulations. Cardinality constrained scheduling imposes the condition that at most $d$0 jobs may be assigned to any machine. In the pure online model on identical machines, the paper shows that a constant competitive algorithm exists, gives a lower bound of $d$1 on the competitive ratio of any online algorithm, and presents a $d$2-competitive algorithm. It also studies ordinal algorithms and a semi-online model with bounded migration, including a transformation from a rate-$d$3 ordinal algorithm to a robust $d$4-approximation with migration factor $d$5 (Epstein et al., 2022). The lower-bound message is that cardinality alone fundamentally changes the online difficulty: the classical identical-machine setting admits sub-2 ratios, whereas the cardinality-constrained version does not (Epstein et al., 2022, Dwibedy et al., 2020).

Another direction couples scheduling with inventory-style costs. In the online joint replenishment problem combined with single-machine scheduling, every unit-time job requires a common resource that must be replenished between its release time and start time at fixed cost $d$6. The objective is

$d$7

The proposed deterministic online algorithm is $d$8-competitive on general input. For $d$9-bounded input, its competitive ratio tends to $\Theta(\log d/\log\log d)$0 as the number of jobs tends to infinity, while lower bounds show that no deterministic algorithm can beat $\Theta(\log d/\log\log d)$1 with two jobs, $\Theta(\log d/\log\log d)$2 with three jobs, or $\Theta(\log d/\log\log d)$3 for sufficiently large $\Theta(\log d/\log\log d)$4-regular instances (Györgyi et al., 2022).

Busy-time scheduling introduces an energy-style cost: a machine with $\Theta(\log d/\log\log d)$5 processors incurs cost whenever it is enabled, independent of whether it is running one job or $\Theta(\log d/\log\log d)$6 jobs. For flexible non-preemptive jobs with release times, deadlines, and processing times, the total objective is machine busy time. In the unbounded-processor case, the paper tightens prior results by proving that no online algorithm can achieve competitive ratio better than $\Theta(\log d/\log\log d)$7 for general flexible jobs, while the best known deterministic upper bound remains $\Theta(\log d/\log\log d)$8. For agreeable jobs, it gives a $\Theta(\log d/\log\log d)$9-competitive online algorithm and a lower bound of $\Theta(\log m+\log d)$0. In the bounded setting, it improves the lookahead-based upper bound from $\Theta(\log m+\log d)$1 to $\Theta(\log m+\log d)$2 using only $\Theta(\log m+\log d)$3 lookahead, and gives the first constant competitive ratio without lookahead in that setting: a deterministic $\Theta(\log m+\log d)$4-competitive algorithm for uniform jobs (Albers et al., 2024).

There is also a practical workflow-oriented line that is explicitly heuristic rather than competitive-analytic. A generalized parallel machine scheduling problem models each job as blocks of tasks with precedence constraints across blocks, and machine-dependent processing times $\Theta(\log m+\log d)$5 that may render a task type infeasible on some machines. The problem is strictly online: when a task is read, it must be scheduled immediately, without knowledge of future tasks, future jobs, or even the remaining blocks of the current job. The paper studies three greedy heuristics—Soonest Finish Time, Soonest Start Time, and Shortest Processing Time—and reports that Soonest Finish Time gives the best makespan in almost all experiments, while Shortest Processing Time is much faster computationally; however, no formal competitive-ratio analysis is provided (Szalkai et al., 2015). This line is useful as a reminder that not all generalized online scheduling work is worst-case competitive theory.

6. Deferred decisions, predictions, learning, and strategic behavior

Recent work expands the model not only by changing the objective or constraint set, but by altering the information structure available to the online scheduler. The online Minimum Peak Appointment Scheduling problem separates online and offline commitments: each arriving request must immediately be assigned a time interval, but the assignment to a room can be deferred to a later offline phase. This deferred decision-making produces a randomized asymptotic competitive ratio under $\Theta(\log m+\log d)$6, improving upon a previous $\Theta(\log m+\log d)$7 bound, while the first known lower bound of $\Theta(\log m+\log d)$8 applies to both deterministic and randomized algorithms (Smedira et al., 2021). The paper explicitly interprets this as evidence that deferring non-essential commitments can improve worst-case performance relative to classical online bin packing.

Prediction-augmented interval scheduling studies a different informational extension: a predicted set of requests is given, possibly with false positives and false negatives. For interval scheduling, the TrustGreedy algorithm achieves

$\Theta(\log m+\log d)$9

and this guarantee is tight. For the more general disjoint path allocation problem, the natural “trust the predicted optimum” strategy achieves a $\min \sum_{S} w_S C_S$0 guarantee, which is optimal for graphs containing an $\min \sum_{S} w_S C_S$1 star subgraph. The same work also characterizes asymptotically tight consistency–robustness trade-offs for randomized interval-scheduling algorithms (Boyar et al., 2023). In a distributed streaming context, POTUS uses predictive service and Lyapunov drift-plus-penalty minimization to reduce response time while guaranteeing queue stability; in the no-prediction case, it proves near-optimal average communication cost within $\min \sum_{S} w_S C_S$2 and backlog bounds that scale linearly with $\min \sum_{S} w_S C_S$3 (Huang et al., 2020).

Learning enters the field in both mechanism-design and direct-optimization forms. In online scheduling with selfish clients and prompt truthful mechanisms, an online mechanism can compete with the best mechanism in a given truthful family by treating mechanisms as experts and managing the cost of switching between them. The clairvoyant regret bound is $\min \sum_{S} w_S C_S$4, and in the non-clairvoyant setting the mechanism achieves $\min \sum_{S} w_S C_S$5 regret, matching lower bounds up to polylogarithmic factors (Chawla et al., 2017). For single-machine min-sum problems with non-decreasing objectives, a unified deep-learning framework uses a time-indexed formulation, offline training on specially generated instances with tractable optimal labels, and online single-instance fine-tuning via a differentiable feasibility surrogate; the reported numerical results cover instances with up to $\min \sum_{S} w_S C_S$6 jobs (Liu et al., 8 Jan 2025).

These developments sit naturally beside the broader trends identified in the makespan survey: resource augmentation, semi-online information, and alternative performance measures are increasingly central to the subject (Dwibedy et al., 2020). A plausible implication is that the modern “online generalized scheduling problem” is no longer defined solely by irrevocable dispatch under adversarial arrivals. It is equally characterized by how much structure is exposed—through group objectives, submodular feasibility, deferred commitments, predictions, speed augmentation, migration, or learning—and by which benchmark is taken as primary: competitive ratio, relative throughput, regret, stability, or empirical approximation quality.