Tela: A Temporal Load-Aware Cloud Virtual Disk Placement Scheme

A cloud provider must decide which physical storage warehouse each customer's virtual disk lives on, and prior schemes place them by average load — but a disk's real load rises and falls over the day, so packing several disks whose busy hours coincide onto one warehouse overloads it even when the averages looked fine; this paper (Huazhong University of Science and Technology, with Tencent — Difan Tan, Ke Zhou, and colleagues) presents Tela, which uses interpretable models to predict each disk's load pattern over time and each warehouse's future peaks, then places disks with complementary rhythms together to shave peaks and balance load — cutting overload occurrences by up to 93.8 percent.

The argument in one glance

Problem

Which storage warehouse a virtual disk lives on decides whether it overloads

Cloud providers sell virtual disks — Cloud Virtual Disks (CVDs) — that appear to a customer's virtual machine as an ordinary block device but are actually backed by big distributed storage clusters the paper calls warehouses. When a customer subscribes to a disk, the provider must decide which physical warehouse to place it on. That placement decision matters a lot: it determines how evenly load is spread across the fleet, how fully each warehouse's capacity is used, and how good the I/O performance is. The natural way to place disks is to look at how much load each one puts on storage — its I/O demand — and pack disks into warehouses so no warehouse exceeds its capacity, balancing the total. But there is a subtlety that prior schemes missed: a disk's load is not a single fixed number. It rises and falls over time, often with a daily rhythm — a disk backing a business application is busy during working hours, a disk holding backups is busy at night. If you place disks by their average load and several disks that all happen to peak at the same time of day end up on the same warehouse, their peaks stack up and the warehouse overloads during those hours, even though its average utilization looked perfectly safe. Ignoring time turns a well-balanced-on-paper placement into one that suffers imbalanced loads, wasted capacity, and frequent, painful overloads.
What people did

Place disks by their load so warehouses stay within capacity

The established approach to CVD placement treats it as a capacity-packing problem: estimate each disk's load, and assign disks to warehouses so that the summed load in each warehouse stays under its limit while spreading load as evenly as possible across the fleet. This is the same instinct used across resource placement — pack items into bins without exceeding each bin's capacity, and balance the totals — applied to virtual disks and storage warehouses, using a load figure for each disk to drive the decision.
Why it fell short

A single load number hides the peaks that actually cause overloads

Summarizing a disk by one load number throws away exactly the information that determines whether a warehouse overloads: when its load happens. Two disks with identical average load can behave completely differently — one steady all day, another idle for twenty hours and then slammed for four — and a placement that only sees the averages cannot tell them apart. So it happily co-locates several disks that all peak in the same window, and the warehouse is fine on average but overloaded during that window, causing I/O latency spikes and violated service levels precisely when customers need performance most. The same blindness wastes capacity in the other direction: to avoid those overloads, operators must leave large safety margins, provisioning warehouses for worst-case peaks that rarely all coincide, so utilization stays low. And because the load is never modeled over time, the scheme cannot deliberately do the one thing that would help — combine disks whose busy periods fall at different times so their peaks fill each other's valleys. What was missing was a placement scheme that understood each disk's load as a pattern over time, and used that to keep warehouse peaks in check.
This paper's idea

Predict each disk's load pattern over time, and place complementary patterns together

Tela is, per the paper, the first temporal load-aware CVD placement scheme. Instead of a single load figure, it treats a disk's load as something that varies over time and predicts that behavior. Using a series of interpretable models — deliberately simple and explainable rather than black-box — Tela predicts the temporal load characteristics and values of each CVD (what its load-over-time pattern looks like and how big it is) and the potential peak loads that would result in each warehouse. Guided by these predictions, Tela places disks into warehouses according to their load patterns, aiming for two things at once: peak shaving and load balancing, while preventing overloads. Peak shaving is the key idea — rather than packing disks that all peak together, Tela groups disks whose busy periods fall at complementary times, so one disk's peak lands in another's quiet stretch and their combined load stays flatter and lower than the sum of their individual peaks. This lets a warehouse safely hold more disks without its aggregate load ever spiking past capacity. Because the models are interpretable, operators can understand and trust why disks are placed as they are. The paper reports that Tela substantially beats the state-of-the-art scheme, reducing overload occurrences by 86.8 to 93.8 percent, cutting the 99th-percentile overload duration by 92.6 percent, and lowering load imbalance by 36.7 to 44.4 percent.

Why this problem is worth solving

Placement is a decision made rarely but lived with constantly, so getting it wrong quietly costs performance and money across the whole fleet for a long time.

When a customer subscribes to a cloud virtual disk, the provider picks a warehouse for it, and the disk tends to stay there. That single decision then governs the disk’s I/O performance and its contribution to a warehouse’s load for the disk’s entire lifetime, which can be months or years. Multiply that across the millions of disks a large cloud runs, and placement quality becomes a dominant factor in how well the whole storage fleet performs and how efficiently its capacity is used. A poor placement is not a transient glitch — it is a standing liability that recurs every day the involved disks are busy together. Two costs follow directly. Overloads hurt customers: when a warehouse’s aggregate load exceeds its capacity, every disk on it suffers latency spikes and possibly violated service-level agreements, damaging exactly the customers whose workloads are heaviest. And the fear of overloads wastes money: to stay safe, operators under-pack warehouses, leaving capacity idle as a buffer against peaks that may never all align. Because the decision is durable, high-volume, and directly tied to both customer experience and provider cost, improving placement pays off broadly and persistently — which is why doing it well, rather than approximately, is worth real effort.

The core realization is that the quantity that governs overloads is not average load but the shape of load over time, and modeling that shape unlocks a strictly better placement.

A warehouse overloads at a moment in time, when the instantaneous sum of its disks’ loads exceeds capacity — not when their averages do. This means the relevant property of a disk is its load curve over time, and the relevant property of a placement is whether, at every instant, the curves stacked in each warehouse stay under the line. Averages are a lossy summary that discards the timing, and timing is precisely what decides overloads, so any scheme built on averages is optimizing the wrong quantity. Once you model the full pattern, a powerful opportunity appears that averages hide: disks with complementary patterns can be combined so their peaks do not coincide. A daytime-busy disk and a nighttime-busy disk placed together produce an aggregate curve much flatter than either alone, so the warehouse can hold both plus more without ever spiking — this is peak shaving, and it is invisible unless you know when each disk is busy. Modeling temporal patterns therefore does two jobs: it prevents the coincidental peak-stacking that causes overloads, and it enables the deliberate peak-shaving that raises how densely warehouses can be packed. Choosing interpretable models to do this matters in production too, because operators must trust and debug placement decisions that affect real customers. Getting the modeled quantity right — patterns, not points — is what makes the whole approach work.

What people did before, and where it broke

Capacity-based CVD placement. Estimate each disk’s load and pack disks into warehouses so summed load stays under capacity, balancing totals. Where it broke. A single load figure ignores when load occurs, so disks that peak simultaneously get co-located and the warehouse overloads during those peaks despite a safe average.
Average-load balancing. Spread disks so average utilization is even across warehouses. Where it broke. Even average utilization can hide synchronized peaks; balancing means at every instant, which averages cannot capture.
Large safety margins to avoid overloads. Under-pack warehouses, provisioning for worst-case peaks. Where it broke. It wastes capacity and lowers utilization by reserving buffer against peaks that rarely all align, and still does not deliberately avoid alignment.
Ignoring complementary patterns. Place disks without regard to the timing of their busy periods. Where it broke. It forgoes peak shaving entirely — the chance to combine disks whose peaks fall at different times so they fill each other’s valleys.
The gap. No scheme modeled each disk’s load as a pattern over time and used those patterns to both prevent coincident peaks and exploit complementary ones, so placement stayed overload-prone and under-utilized.

How Tela works

From average-load packing to temporal-pattern-aware placement

↓ →
new
↓ →
new
↓ →
new
↓ →
new
↓ →
new

The idea

Predict future behavior with simple, explainable models to guide the decision. Tela forecasts each disk’s load over time and each warehouse’s peaks, and drives placement from those predictions rather than from a static snapshot. Building a prediction of how resources will behave and using it to make better systems decisions ahead of time is the same forecasting approach the work on a general prediction service for cloud systems (the Prediction System Service) covered elsewhere on this site is built on, turning anticipated behavior into better provisioning. Tela adds the discipline of keeping the models interpretable, so operators can trust placements that affect real customers.

Pack items into limited capacity while respecting constraints over time. At heart, placing disks into warehouses is a packing problem — fit as much as possible into each bin without violating its limits — made harder because the limit must hold at every instant. Assigning items to fixed-capacity resources so that constraints are satisfied and space is used densely is the same allocation-and-packing challenge the work on efficient on-chip memory allocation for production machine learning (TelaMalloc) covered elsewhere on this site tackles, fitting demands into limited capacity without overrunning it. Tela solves this packing over the time dimension, where the constraint is a peak that must never be exceeded at any moment.

Control I/O load so no shared resource is overwhelmed. Tela’s goal is to keep any warehouse from being swamped by the combined I/O of the disks it hosts, protecting performance for everyone on it. Managing and shaping the I/O load placed on a shared storage resource so that contention does not overwhelm it is the same I/O-load-control concern the work on block I/O control for containers in datacenters (iocost) covered elsewhere on this site addresses, keeping a shared storage backend from being saturated. Tela pursues that goal at placement time, deciding what shares a warehouse so its I/O never spikes past capacity, rather than throttling after the fact.

Following a placement decision through Tela

Placing one new virtual disk without stacking peaks

An illustration, at the level the paper's abstract describes, of how Tela places a Cloud Virtual Disk. The full paper was not available for this explainer, so internal mechanism details are omitted.

Step 1 of 5

A new disk needs a warehouse

A customer subscribes to a Cloud Virtual Disk, and Tela must choose which physical storage warehouse to place it on, aiming to avoid overloads and keep the fleet balanced.

The choice will affect this disk and its warehouse for a long time.

Why it works

Modeling patterns catches the peaks averages hide. A single load number cannot show when a disk is busy → but predicting its load-over-time pattern → so Tela sees the peaks that actually cause warehouse overloads.
Predicting warehouse peaks checks the right limit. Capacity is exceeded at moments, not on average → but forecasting each warehouse’s peak load over time → so placement is validated against the constraint that truly matters.
Complementary placement shaves peaks. Co-locating disks that peak together stacks their loads → but grouping disks whose busy periods fall at different times → so their combined curve stays flatter and lower than the sum of peaks.
Peak shaving raises safe density. Fear of overloads forces wasteful safety margins → but keeping combined peaks low by design → so warehouses hold more disks at higher utilization without spiking.
Interpretable models earn operator trust. Black-box placement is hard to rely on in production → but using simple, explainable models → so operators can understand and trust why each disk is placed where it is.

What it costs (trade-offs and scope)

Predictions can be wrong. Tela relies on forecasting each disk’s temporal load; a disk whose behavior shifts unexpectedly, or a new disk with little history, may be mispredicted, and the quality of placement depends on the models’ accuracy.
Interpretable models trade some accuracy for trust. Choosing simple, explainable models over complex black-box ones is deliberate, but such models may capture less nuance than a heavier learned approach, a trade the paper makes for production trustworthiness.
It targets placement, not the whole lifecycle. Tela decides where to place a disk; workloads that drift over time might eventually benefit from migration, which is a separate mechanism the placement decision alone does not perform.
Benefits depend on patterns being complementary and predictable. Peak shaving works when disks have distinct, stable busy periods that can be paired; a fleet where every disk peaks at the same time offers less room to shave.
This explainer is built from the abstract only. The specific interpretable models, how patterns and peaks are predicted, the placement algorithm, the evaluation setup, and the conditions behind the reported reductions are in the full paper, which was not available here; only the abstract’s stated design and headline numbers are reported.

The comparison

Placing cloud virtual disks onto storage warehouses

Property	Capacity-based placement	Average-load balancing	Tela
What load information it uses
Checks the right limit
Exploits complementary patterns
Effect on utilization
Reported outcome

A comparison of ways to place cloud virtual disks. Capacity-based and average-load approaches summarize each disk by a single load figure and so co-locate disks whose peaks coincide, causing overloads and forcing wasteful safety margins. Tela predicts each disk's load pattern over time and each warehouse's future peaks with interpretable models, then places disks with complementary rhythms together to shave peaks and balance load — reducing overload occurrences by 86.8 to 93.8 percent, cutting 99th-percentile overload duration by 92.6 percent, and lowering load imbalance by 36.7 to 44.4 percent versus the state of the art. Exact figures should be read from the paper.Underlined values explain themselves when clicked.

Where this stops working

What to keep in mind

This explainer is based only on the paper’s abstract and TLDR (obtained via Semantic Scholar); the full text was not available — the paper is behind the ACM Digital Library with no open preprint reachable at the time of writing, so this page is marked a draft. What is reported here is what the abstract states: Cloud Block Storage relies on Cloud Virtual Disks to give block interfaces to cloud VMs, and placing those disks onto physical storage warehouses affects resource utilization, load balancing, and I/O performance; prior works ignored temporal fluctuations in load, causing imbalance, low utilization, and frequent overloads. Tela is described as the first temporal load-aware CVD placement scheme: it uses a series of interpretable models to predict the temporal load characteristics and values of CVDs and potential warehouse peak loads, and places CVDs according to their load patterns for peak shaving and load balancing while preventing overloads. The abstract reports that Tela reduces overload occurrences by 86.8 to 93.8 percent, cuts 99th-percentile overload duration by 92.6 percent, and decreases load imbalance by 36.7 to 44.4 percent versus the state-of-the-art scheme. Everything beyond these claims — the specific models, the prediction and placement algorithms, the evaluation data and system, and the detailed results — is not contained in the abstract and should be read in the full paper. The work is associated with production traces released by Tencent. Background on cloud block storage, virtual disks, load balancing, temporal workload patterns, and peak shaving is standard context, and the step-by-step walkthrough is an illustration at the abstract’s level, not a verbatim trace of the paper’s mechanism.

Background

Why cloud resource management keeps rediscovering the time dimension

It helps to see Tela within the broad effort to manage cloud resources efficiently. Providers run enormous fleets of shared hardware, and the recurring question is how to pack many tenants’ demands onto that hardware without anyone suffering, while keeping utilization high. A naive answer treats each demand as a fixed size and solves a static packing problem, but real workloads are not static — they ebb and flow, often with strong daily rhythms — and the systems that win are usually the ones that take this temporal structure seriously. The pattern shows up everywhere: harvesting resources that are idle now but needed later, scheduling around predicted demand rather than reacting to it, and combining workloads whose busy periods complement each other so shared capacity is used smoothly. The recurring lesson is that averages lie about contention, which happens at peaks, so modeling behavior over time is what separates placements that overload from placements that do not. This drive to use shared cloud infrastructure densely and safely by understanding how demand varies over time is the same terrain the work on disaggregated storage in modern datacenters (disaggregated RAID storage) covered elsewhere on this site occupies, organizing storage resources for efficient, reliable sharing at scale. Tela’s particular contribution is to bring temporal awareness to virtual disk placement specifically, and to show, on production-scale data, that predicting load patterns and shaving peaks dramatically cuts the overloads that static placement causes.

If you remember only three things

1. Cloud providers place customers' virtual disks (Cloud Virtual Disks) onto physical storage warehouses, and that placement decision — durable and made across millions of disks — governs load balancing, capacity utilization, and I/O performance. Prior schemes summarize each disk by a single load figure and pack disks so warehouses stay under capacity on average. But a disk's load is not constant: it rises and falls over time, often with a daily rhythm. When disks that happen to peak at the same time of day are co-located, their peaks stack and the warehouse overloads during those hours even though its average looked safe — causing latency spikes and SLA violations — and the fear of this forces operators to under-pack warehouses, wasting capacity.
2. Tela is, per the paper, the first temporal load-aware CVD placement scheme. Rather than a single number, it treats a disk's load as a pattern over time and predicts it: using a series of interpretable (deliberately simple and explainable) models, it forecasts each disk's temporal load characteristics and values and the potential peak load each warehouse would reach. Guided by these predictions, it places disks according to their patterns to achieve peak shaving and load balancing together while preventing overloads. Peak shaving is the crux: instead of packing disks that peak together, Tela pairs disks whose busy periods fall at complementary times, so one disk's peak lands in another's valley and their combined load stays flatter and lower than the sum of peaks — letting a warehouse hold more disks without ever spiking past capacity.
3. Because overloads happen at moments, not on average, modeling load over time is optimizing the right quantity, and it both prevents coincidental peak-stacking and enables deliberate peak-shaving that raises safe density; interpretable models let operators trust the placements. The paper reports that Tela substantially beats the state-of-the-art scheme, reducing overload occurrences by 86.8 to 93.8 percent, cutting 99th-percentile overload duration by 92.6 percent, and lowering load imbalance by 36.7 to 44.4 percent. The trade-offs: predictions can be wrong (especially for new or drifting disks), interpretable models trade some accuracy for trust, it addresses placement rather than later migration, and peak shaving needs disks with distinct, predictable, complementary rhythms. This explainer is drawn only from the abstract, so the models and algorithms should be read in the full paper.

Check yourself

Check your understanding

1. Why does ignoring temporal load fluctuations cause warehouse overloads in virtual disk placement?
2. What does Tela predict, and why interpretable models?
3. What is peak shaving, and how does Tela use it?
4. What did Tela achieve, and how should this page be read?

Sources

The paper is “Tela: A Temporal Load-Aware Cloud Virtual Disk Placement Scheme” by Difan Tan, Jiawei Li, Hua Wang, Xiaoxiao Li, Wenbo Liu, Zijin Qin, Ke Zhou, Ming Xie, and Mengling Tao (Huazhong University of Science and Technology, with industry collaboration; associated production traces have been released by Tencent), published at ASPLOS 2025, DOI 10.1145/3669940.3707252.

This explainer is written only from the paper’s abstract and TLDR, obtained via the Semantic Scholar API; the full text was not available, because the paper is behind the ACM Digital Library and no open preprint was reachable at the time of writing. Accordingly this page is marked a draft. The abstract describes Tela as the first temporal load-aware Cloud Virtual Disk placement scheme: it uses a series of interpretable models to predict the temporal load characteristics and values of CVDs and potential warehouse peak loads, then places CVDs according to their load patterns for peak shaving and load balancing while preventing overloads. The abstract reports that Tela reduces overload occurrences by 86.8 to 93.8 percent, cuts 99th-percentile overload duration by 92.6 percent, and decreases load imbalance by 36.7 to 44.4 percent versus the state-of-the-art scheme. Background on cloud block storage, cloud virtual disks, placement and load balancing, temporal workload patterns, and peak shaving is standard, widely-known context. All quantitative claims are the paper’s own; the internal mechanisms and full evaluation should be confirmed in the paper.