Dynamic Vector Bin Packing (DVBP)

• Dynamic Vector Bin Packing is a framework for online scheduling of multi-dimensional, time-varying tasks that ensures concurrent resource constraints on homogeneous servers.
• Its methodology employs Any Fit algorithms such as First Fit, Next Fit, and Move To Front, with competitive analysis providing insights through tight upper and lower bounds.
• Practical data reduction techniques like time-compression and (1+ε)-deletion enable significant instance size shrinkage while preserving near-optimal resource allocation.

Dynamic Vector Bin Packing (DVBP) generalizes classical bin packing by capturing the temporal and multi-dimensional nature of resource allocation in cloud computing scenarios, where tasks (jobs) exhibit vector-valued resource requirements and overlapping lifespans. The DVBP problem models online assignment of such jobs to homogeneous servers (bins), striving to minimize the total usage or number of bins while respecting concurrent multi-resource constraints at all times. This framework encompasses both practical scheduling applications and a rich theoretical landscape spanning competitive analysis, non-clairvoyant scheduling, and parameterized complexity.

1. Formal Definition and Problem Variants

DVBP is defined by a set of $n$ jobs (requests) %%%%1%%%%, each with:

• Arrival time $a(r)\in\mathbb Q_{\ge0}$
• Departure time $e(r)\in\mathbb Q_{\ge0}$, with lifespan $I(r) = [a(r),e(r))$
• $d$-dimensional resource demand $s(r)\in[0,1]^d$ (or, in the integer variant, $a^{(i)}\in\mathbb N^d$)

At any time $t$, an assignment of jobs to bins $B_1,\dots,B_k$ must guarantee

$$\sum_{r\in B_j, a(r)\le t<e(r)} s(r) \leq \mathbf{1}$$

(coordinate-wise), where $\mathbf{1}=(1,\ldots,1)$ represents the per-bin capacity.

Objective functions in DVBP differ according to the underlying system model:

• MinUsageTime DVBP: Minimize total server rental time,

$$\mathrm{cost}(A,R) = \sum_{i=1}^m \operatorname{span}(R_i)$$

where $\operatorname{span}(R_i)$ denotes the union of activity intervals of jobs in bin $B_i$. The competitive ratio is

$$\mathrm{CR}(A) = \sup_R \frac{\mathrm{cost}(A,R)}{\mathrm{OPT}(R)}$$

with $\mathrm{OPT}(R)$ defined as the optimal (possibly clairvoyant) solution.

• Classical (static) DVBP: Minimize the total number of bins required over the interval set, subject to the multi-dimensional packing constraints at every time.

Parameterization frequently involves:

• $d$: number of resource dimensions
• $\mu = \max_r \ell(r)$ ($\ell(r) = e(r) - a(r)$): duration ratio of the longest to shortest job
• $h$: maximum concurrent jobs (instance "height")
• $\varphi$: number of distinct job "flavors" (unique demand vectors)

2. Any-Fit Algorithms and Algorithmic Schemes

Any Fit algorithms are a class of online procedures that open new bins only if existing bins cannot accommodate an incoming job $r$ without exceeding the resource constraint in any dimension. Core representatives include:

• First Fit (FF): Maintain bins in order of creation; scan from oldest to newest to find the first feasible bin for $r$.
• Next Fit (NF): Only one bin is kept open at any time; if $r$ does not fit, close the current bin and open a new one.
• Move To Front (MTF): Bins are ordered by recent use; an assigned bin moves to the front, biasing toward temporal alignment among jobs.
• Best Fit (BF): Place $r$ in the bin that would maximize post-assignment bin utilization (largest $\|L(B)+s(r)\|_{\infty}$ without overflow); known to have unbounded worst-case ratio even for $d=1$.

Upon job departure, bins emptied of all jobs are removed from tracking.

The schematic pseudocode for Any Fit algorithms is:

Initialize empty bin list L. For each arriving job r: Scan L in prescribed order for bin B where L(B) + s(r) <= 1 (coordinatewise). If found, assign r to B (and update L order as required). Else, open new bin B', add to L, assign r to B'. Upon job departure: If a bin becomes empty, remove it from L.

The specific reordering and scanning protocols distinguish algorithm variants and critically impact competitive behavior.

3. Competitive Analysis: Upper and Lower Bounds

The theoretical performance of Any Fit algorithms is captured via worst-case competitive ratio bounds, dependent on $\mu$ and $d$:

Proven Lower Bounds

• Any Fit (all $d$): $\mathrm{CR}\ge (\mu+1)d$
• Next Fit: $\mathrm{CR}\ge 2\mu d$
• Move To Front ($d=1$): $\mathrm{CR}\ge 2\mu$

These bounds are established via adversarial instance constructions in which carefully crafted job sequences force inefficiencies in online assignment relative to an optimal clairvoyant scheduler.

Nearly Tight Upper Bounds

The main results for MinUsageTime DVBP yield:

Algorithm	Competitive Ratio Upper Bound
Move To Front	$(2\mu+1)d+1$
First Fit	$(\mu+2)d+1$
Next Fit	$2\mu d+1$

The Move To Front analysis partitions each bin's active interval into leading and non-leading segments, showing that non-leading intervals can be bounded in aggregate by $(2\mu+1)d\cdot \mathrm{OPT}$, while leading intervals sum to at most $\mathrm{OPT}$. The gap between lower and upper bounds for MTF in $d\geq 2$ remains open.

First Fit and Next Fit bounds utilize dimension-aware charge arguments to account for multi-dimensional demand and temporal overlap.

4. Data Reduction and Approximation in Large-Scale DVBP

The complexity of DVBP instances motivates preprocessing approaches to shrink problem size while retaining (near-)optimality.

Kernelization and Lower Bounds

It is proven that DVBP admits no polynomial-size kernel in $(h+\varphi)$, even for $(1+\varepsilon)$-approximation, unless the polynomial hierarchy collapses. Only an exponential-size kernel of $O(h\cdot 2^h)$ is available.

Practical Polynomial-Time Reductions

Despite theoretical hardness, two data reduction rules achieve dramatic real-world instance shrinkage:

• Time-compression: Remaps all start/end points onto a contiguous range $[1,\dotsc,T']$ with $T'\leq 2n$, preserving all job interval overlaps, often reducing the number of types $T$ by a factor of $\sim 6$.
• $(1+\varepsilon)$-deletion rule: Greedily deletes up to $\lfloor \varepsilon L \rfloor$ bins' worth of requests from the instance, where $L$ is a time-indexed lower bound on the number of bins, enabling any $\alpha$-approximate algorithm on the reduced instance to guarantee an $\alpha (1+\varepsilon)$-approximation overall.

These techniques achieve order-of-magnitude reductions in instance size on industry traces (e.g., Azure, Huawei) while preserving bounded optimality gaps and polynomial running time (Murhekar et al., 2023, Bevern et al., 2022).

5. Experimental Evaluations and Empirical Findings

Extensive experiments on both synthetic and real-world traces illustrate key behaviors:

Online Algorithms:

• On synthetic data ($d\in\{1,2,5\}$, varying $\mu$), Move To Front exhibits the best average-case ratio for all evaluated parameter regimes.
• First Fit and Best Fit perform comparably (with low variance waste), while Next Fit's performance deteriorates with increasing $\mu$.
• Heuristics such as Worst Fit and Random Fit display high packing waste and variability.

Data Reduction:

• Time-compression alone reduces type counts by $\sim 1/3$ without reducing items.
• The combination of time-compression and $(1+\varepsilon)$-deletion ($\varepsilon=0.05$) reduces item counts by $12.6\times$ (Huawei) and $5.4\times$ (Azure), and type counts by $\sim 50\times$.
• Reduced instances retain at least $95\%$ of the theoretical maximum deletions, with overall instance size reduction enabling tractability for stronger solvers.

Dataset	Initial $n$	$n$ after reduction	Initial $\tau$	$\tau$ after reduction	Utilization $R$	Factor $K$
Huawei	111,774	8,849	9	5	0.969	1.576
Azure	3,792,000	704,832	74	49	0.946	1.319

6. Open Questions and Research Directions

Several unresolved topics and potential advancements in DVBP research are identified:

• Closing competitive ratio gaps: For $d\geq 2$, the gap between lower $(\mu+1)d$ and upper $(2\mu+1)d+1$ bounds for Move To Front persists.
• Algorithmic improvements for low dimensions: Specialized designs for small $d$ ($d=2$ or $3$) relevant to practical cloud settings may yield better theoretical and empirical results.
• Beyond worst-case: Best Fit exhibits unbounded worst-case but competitive average-case ratios, suggesting the need for distributional or smoothed analyses.
• Clairvoyant or stochastic enhancements: Leveraging predictions (including machine-learned models) of future arrivals or durations remains an open frontier, especially for $d$-dimensional cases.
• Multiple objectives: Incorporation of real-world criteria such as power usage, migration overheads, or startup latencies suggests multi-objective DVBP formulations.

In summary, DVBP offers a mathematically rigorous abstraction for dynamic, multi-resource online scheduling, with a deep body of work characterizing both worst-case guarantees and practical reduction techniques for large-scale deployments in cloud environments (Murhekar et al., 2023, Bevern et al., 2022).