PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteCloudflare says it reclaimed 100 TB of RAM across its network by changing how its Pingora Backend Router stored and counted consistent-hash points—not by installing different hardware. The transferable lesson is to measure the memory cost of a data structure, model the quality trade-off behind its size, validate the change against real workloads, and deploy with a rollback path. The figures are Cloudflare’s published results, not independently audited benchmarks.
Where Cloudflare found the memory cost
Cloudflare’s September 18, 2026 engineering article describes optimizing Pingora Backend Router (PBR), its internal load-balancing service. PBR routes cacheable requests according to a request’s URL, helping keep a file’s stored copy in a stable location and avoid needless duplication within a data center. The account is in Cloudflare’s engineering article, “Saving another 100TB of RAM with math (and Rust)”.
PBR uses consistent hashing: server identities and request keys are mapped into a shared hash space, and a request is directed to a nearby server point. Compared with a naive mapping, this can limit how many assignments change when servers are added or removed. But one point per server can yield uneven ranges. Multiple points help balance assignments, while server weights can represent differences in storage capacity.
Production requirements can multiply the structure. Weights may call for more points, and compliance rules or cache features may require separate rings for different server subsets. Cloudflare says some instances were using about 6 GB of excessive memory before the changes.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →#1 Best Overall
- Capacity: 16GB (2x 8GB Modules) | Type: DDR3 240-Pin | Speed: 1600MHz PC3-12800 / (PC3-12800E) | ECC Type: ECC-UDIMM (ECC Unbuffered DIMM) | Rank: 2Rx8 (Dual Rank x8) | Voltage: 1.35V
- Designed for ECC UDIMM Compatible Servers/Workstations (Rated Speeds & ECC Capabilities are CPU Dependent). Not Compatible with Desktops/Laptops.
- ECC Types can not be mixed | All installed modules must be ECC UDIMMs in order to function properly | A maximum of eight ranks per memory channel can be installed at once
- All A-Tech memory modules undergo stringent quality control testing to ensure dependable and reliable performance
- Backed by A-Tech's Limited Lifetime Warranty + Tech Support Team available to help before and after your purchase
Two changes reduced the ring’s memory footprint
Pack each point into six bytes
A hash point needs to identify both a position in the hash space and its server. Cloudflare changed its representation from eight bytes to six: a 32-bit hash plus a 16-bit server index, stored in a six-byte array with accessors. The authors chose a 16-bit index because they considered more than 65,000 simultaneously coordinated servers unlikely in this use case.
The compact form matters because a normal Rust struct may be padded to eight bytes to meet alignment requirements, even when its fields total six bytes. Cloudflare reports that the six-byte representation reduced memory used by consistent-hashing storage by 25%. The 16-bit index is a design choice for this environment, not a safe default for systems that may need a larger index space.
Use fewer points where the balance benefit is small
Cloudflare also modeled how the number of points affects load distribution. Its article derives an error expression for k hashes per server using expected value, standard deviation, and coefficient of variation. In the article’s example, increasing the ring from 10,000 to 100,000 points per server reduced the modeled error by only 0.7% for the final 90,000 points. Cloudflare says it cut the points per server by 90% in its revised configuration without appreciable error in its system.
Rank #2
- A-Tech RAM Memory compatible for select DDR4 Server and Workstation systems only; (*WILL NOT WORK with Desktop or Laptop Computers/PCs*)
- 32GB RAM Kit (2 x 16GB Modules); DDR4 DIMM 288 Pin; Speeds up to 2133MHz PC4-17000 (PC4-2133P)
- ECC Unbuffered UDIMM; 2Rx8 - Dual Rank x8; JEDEC DDR4 standard 1.2V
- Improves system performance, workload capacity, and reduces bottlenecks by increasing memory (RAM) resources
- Note: This memory is ECC Unbuffered and cannot be mixed with different ECC types such as ECC Registered, ECC Load Reduced, or Non-ECC Unbuffered; (Memory compatibility can vary among different system models and their installed components; please verify compatibility and follow memory channel guidelines to ensure maximum performance)
That is a case-specific result, not a universal threshold: an acceptable ring size depends on the distribution your service needs and the cost of imbalance. The authors also note that their idealized model assumes a continuous ring, whereas production uses 32-bit hashes. Collisions in that finite hash space can add error, particularly as point counts rise.
How to apply the method to another system
- Find the costly representation. Identify a frequently replicated or large in-memory structure, then measure its actual footprint at the scale where it matters. Include copies, indexes, padding, and variants maintained for different rules or features.
- Separate representation cost from algorithmic quality. Ask whether fields can be packed more tightly, but check alignment and access requirements in the language and runtime you use. Do not choose narrower identifiers unless the maximum number of entities they can represent is safely above your real needs.
- Model the trade-off. Estimate how reducing the structure changes the service’s quality—such as distribution error, collision risk, or lookup behavior. Treat the model as a way to choose candidates for testing, not as a substitute for production measurements.
- Test against representative behavior. Compare memory savings with the operational metric the structure exists to protect. For a routing ring, that means checking assignment balance and the consequences of remapping, not memory alone.
- Deploy in stages with a reversal path. Make the new and old behavior distinguishable, start with limited traffic or locations, monitor relevant service and workload metrics, and keep a practical way to restore the prior behavior.
Why Cloudflare migrated gradually
A changed hash ring can send cacheable requests to different servers. That can reduce cache locality, cause misses, and increase requests to origin systems. Cloudflare therefore did not replace the ring everywhere at once: it temporarily ran both versions, used a migration framework to select a ring for requests, and expanded deployment from small validation locations to progressively larger data-center groups.
The team monitored backend-selection traces, ring-version counters, connection errors, process memory, startup time, cache behavior, and origin traffic. After completing the migration, it removed the old path. The sequence illustrates that the savings depended not only on compact data and fewer points, but also on managing the workload transition safely.
Rank #3
- A-Tech RAM Memory compatible for select DDR5 Server systems; (WILL NOT WORK with Desktop Computers/PCs or Laptop Computers)
- Single 64GB RAM Module; DDR5 DIMM 288 Pin; Speeds up to 5600MHz PC5-44800 (PC5-5600B)
- ECC Registered RDIMM; 2Rx4 (EC8, 10x4) - Dual Rank x4; JEDEC DDR5 standard 1.1V
- Improves system performance, workload capacity, and reduces bottlenecks by increasing memory (RAM) resources
- Note: EC8 (10x4) ECC Registered modules cannot be mixed with EC4 (9x4) ECC Registered modules or with different ECC types such as ECC Unbuffered, ECC Load Reduced or Non-ECC Unbuffered; (Memory compatibility can vary among different system models and their installed components; please verify compatibility and follow memory channel guidelines to ensure maximum performance)
What the result does—and does not—show
Cloudflare’s reported 100 TB is a global outcome for its own fleet and workload, built from two changes to its consistent-hashing structures. It does not mean another service can expect the same savings, or that changing hardware was responsible. The useful precedent is the process: inspect an expensive structure, quantify the trade-off between its footprint and its purpose, verify behavior, and roll out with monitoring and rollback.
Cloudflare says the modified implementation is available in the open-source pingora-ketama crate as an unadvertised Cargo feature. The article’s authors are Kevin Guthrie, Mariia Iurchenko, Zaidoon Abd Al Hadi, and Ivan Babrou.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
