Painful cold boots, terrible auto-scale times, minutes-long waits for compute nodes to be up: these are standard headaches that cloud engineers have to deal with, and work-around, on a daily basis. Much of this stems from the fact that cloud infrastructure relies on technology that was never intende...
Painful cold boots, terrible auto-scale times, minutes-long waits for compute nodes to be up: these are standard headaches that cloud engineers have to deal with, and work-around, on a daily basis. Much of this stems from the fact that cloud infrastructure relies on technology that was never intended to be reactive in millisecond scales, or at least within the round-trip time of a user request.
In this talk, Felipe will show how building an edge platform based on technology *intended* to be fast and reactive can lead to substantial gains in both the efficiency of running the platform and the user experience when using it. First, Felipe will demonstrate how leveraging unikernels built with the Unikraft
Linux Foundation project can lead to (cold) boot times of as little as 2 ms. This, coupled with a fast VMM (FireCracker), and the fact that the images consume little memory (a few MBs) means that 10-50K of them can be run on a single, inexpensive server. Even better: fast suspend/resume times mean that instances can be suspended when idle, and then resumed just-in-time, as the next request comes in — all transparent to the user.
Felipe will show extensive benchmarks (cold boot times, latency, throughput, memory consumption, etc.) from such technology to give a taste of what future cloud platforms could (and should) look like. Felipe will wrap up the talk by showing how Unikraft integrates with the wider tooling ecosystem (e.g., Kubernetes).
Size: 5.98 MB
Language: en
Added: Jun 26, 2024
Slides: 56 pages
Slide Content
Unleashing 10x More Efficient Cloud Deployments with KraftCloud Felipe Huici CEO & Co-Founder at Unikraft
Felipe Huici CEO & Co-Founder at Unikraft Something cool I’ve done : building a revolutionary millisecond cloud platform 😀 My perspective on P99s : performance is fundamental! Another thing about me : from Argentina What I do away from work : sports, guitar, reading, photography
2.6GB
2.6GB
2.6GB
2.6GB 1.9MB
With Linux With Unikraft
HOW? Specialization
APP
APP
APP LIBRARIES
APP LIBRARIES OS/KERNEL
APP LIBRARIES OS/KERNEL
APP LIBRARIES OS/KERNEL SPECIALIZED VM UNIKERNEL
Unikraft Three Principles Fully Modular - allow for full specialization Linux API Compatibility - run unmodified apps and langs Tooling Integration - keep using the tools you love
Unikraft Arch
The Millisecond Cloud Platform
c Linux Host (KVM) BAREMETAL SERVER KraftCloud Arch
c Linux Host (KVM) BAREMETAL SERVER KraftCloud Arch
c Linux Host (KVM) ukp REST API BAREMETAL SERVER KraftCloud Arch
c Linux Host (KVM) ukp REST API BAREMETAL SERVER KraftCloud Arch image registry
c Linux Host (KVM) ukp u1 REST API BAREMETAL SERVER KraftCloud Arch image registry
c Linux Host (KVM) ukp u1 u2 REST API BAREMETAL SERVER KraftCloud Arch image registry
c Linux Host (KVM) ukp u1 u2 REST API u3 BAREMETAL SERVER KraftCloud Arch image registry
c Linux Host (KVM) ukp u1 u2 REST API u3 … u10k BAREMETAL SERVER KraftCloud Arch image registry
KraftCloud Differentiators Scale to zero
KraftCloud Differentiators Scale to zero Low price point
KraftCloud Differentiators Scale to zero Low price point Deploy to 70+ edge locations
KraftCloud Differentiators Scale to zero Low price point Deploy to 70+ edge locations Tiny images
KraftCloud Differentiators Scale to zero Low price point Deploy to 70+ edge locations Tiny images Death to cold boots
KraftCloud Differentiators Scale to zero Low price point Deploy to 70+ edge locations Tiny images Death to cold boots Immediate autoscale
KraftCloud Differentiators Scale to zero Low price point Deploy to 70+ edge locations Tiny images Death to cold boots Immediate autoscale Full HW-isolation, secure and minimal TCB
KraftCloud Differentiators Scale to zero Low price point Deploy to 70+ edge locations Tiny images Death to cold boots Immediate autoscale Full HW-isolation, secure and minimal TCB Reduced carbon footprint
KraftCloud in Practice Sign up for free at https://kraft.cloud/ 😀 Receive an account name and token Follow guides on https://docs.kraft.cloud/ To interact with platform: REST API kraft tool Docker integration Kubernetes integration Join our Discord servers OSS, see https://unikraft.org/ (Linux Foundation open source project) KraftCloud (for tech support, after sign up)
Cold Boot Times vs Linux Application Unikraft Linux NGINX 4.2 ms 715 ms Redis 7.8 ms 761 ms SQLite 5.2 ms 698 ms Nodejs 47 ms 820 ms Go HTTP server 8.8 ms 688 ms VMM: Firecracker
Security - TCB Reduction (NGINX) Linux Unikraft Image Size 30.4MiB 1.9MBs/1.25 MiB (DCE) Lines of Code (kernel + app) 5,501,759 283,622 Number of Symbols 192,969 4,933
Guest Latency (Unikraft)
Throughput - Unikraft vs Linux (NGINX) Linux: Ubuntu 22.04.3 LTS (Linux 6.5 with firecracker configuration, nginx 1.18.0), VM with 1GiB memory
Max Density (NGINX Guests) hw specs : 2x Intel Xeon 6338, 64 cores @ 2.0 GHz, 1TiB memory), Unikraft VM with 12MiB memory