"lightweight virtualization with linux containers and docker". jerome petazzoni, dotcloud
Post on 28-Jan-2015
110 Views
Preview:
DESCRIPTION
TRANSCRIPT
привет!
меня зовут ДжеромЯ не говорю по России
(к сожалению)
Lightweight Virtualizationwith
Linux Containersand
Docker
Yet another Conference – Moscow, 2013
Jérôme Petazzoni, dotCloud Inc.
Яндекс:спасибо большое!
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
Why Linux Containers?
What arewe tryingto solve?
The Matrix From Hell
The Matrix From Hell
Many payloads
● backend services (API)● databases● distributed stores● webapps
Many payloads
● Go● Java● Node.js● PHP● Python● Ruby● …
Many payloads
● CherryPy● Django● Flask● Plone● ...
Many payloads
● Apache● Gunicorn● uWSGI● ...
Many payloads
+ your code
Many targets
● your local development environment● your coworkers' developement environment● your Q&A team's test environment● some random demo/test server● the staging server(s)● the production server(s)● bare metal● virtual machines● shared hosting
+ your dog's Raspberry Pi
Many targets
● BSD● Linux● OS X● Windows
Many targets
● BSD● Linux● OS X● Windows
Not yet
The Matrix From Hell
Static website ? ? ? ? ? ? ?
Web frontend ? ? ? ? ? ? ?
background workers ? ? ? ? ? ? ?
User DB ? ? ? ? ? ? ?
Analytics DB ? ? ? ? ? ? ?
Queue ? ? ? ? ? ? ?
Development VM QA Server Single Prod Server Onsite Cluster Public Cloud Contributor’s laptop Customer Servers
Real-world analogy:containers
Many products
● clothes● electronics● raw materials● wine● …
Many transportation methods
● ships● trains● trucks● ...
Another matrix from hell
? ? ? ? ? ? ?
? ? ? ? ? ? ?
? ? ? ? ? ? ?
? ? ? ? ? ? ?
? ? ? ? ? ? ?
? ? ? ? ? ? ?
Solution to the transport problem:the intermodal shipping container
Solution to the transport problem: the intermodal shipping container
● 90% of all cargo now shipped in a standard container● faster and cheaper to load and unload on ships
(by an order of magnitude)● less theft, less damage● freight cost used to be >25% of final goods cost, now <3%● 5000 ships deliver 200M containers per year
Solution to the deployment problem: the Linux container
Linux containers...
● run everywhere– regardless of kernel version
– regardless of host distro
– (but container and host architecture must match*)
● run anything– if it can run on the host, it can run in the container
– i.e., if it can run on a Linux kernel, it can run
*Unless you emulate CPU with qemu and binfmt
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
What are Linux Containers exactly?
High level approach:it's a lightweight VM
● own process space● own network interface● can run stuff as root● can have its own /sbin/init
(different from the host)
« Machine Container »
Low level approach:it's chroot on steroids
● can also not have its own /sbin/init● container = isolated process(es)● share kernel with host● no device emulation (neither HVM nor PV)
« Application Container »
Separation of concerns:Dmitry the Developer
● inside my container:– my code
– my libraries
– my package manager
– my app
– my data
Separation of concerns:Oleg the Ops guy
● outside the container:– logging
– remote access
– network configuration
– monitoring
How does it work?Isolation with namespaces
● pid● mnt● net● uts● ipc● user
pid namespace
jpetazzo@tarrasque:~$ ps aux | wc -l212
jpetazzo@tarrasque:~$ sudo docker run -t -i ubuntu bashroot@ea319b8ac416:/# ps auxUSER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMANDroot 1 0.0 0.0 18044 1956 ? S 02:54 0:00 bashroot 16 0.0 0.0 15276 1136 ? R+ 02:55 0:00 ps aux
(That's 2 processes)
mnt namespace
jpetazzo@tarrasque:~$ wc -l /proc/mounts
32 /proc/mounts
root@ea319b8ac416:/# wc -l /proc/mounts
10 /proc/mounts
net namespace
root@ea319b8ac416:/# ip addr
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00 inet 127.0.0.1/8 scope host lo valid_lft forever preferred_lft forever inet6 ::1/128 scope host valid_lft forever preferred_lft forever
22: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP qlen 1000 link/ether 2a:d1:4b:7e:bf:b5 brd ff:ff:ff:ff:ff:ff inet 10.1.1.3/24 brd 10.1.1.255 scope global eth0 valid_lft forever preferred_lft forever inet6 fe80::28d1:4bff:fe7e:bfb5/64 scope link valid_lft forever preferred_lft forever
uts namespace
jpetazzo@tarrasque:~$ hostnametarrasque
root@ea319b8ac416:/# hostnameea319b8ac416
ipc namespace
jpetazzo@tarrasque:~$ ipcs------ Shared Memory Segments --------key shmid owner perms bytes nattch status0x00000000 3178496 jpetazzo 600 393216 2 dest 0x00000000 557057 jpetazzo 777 2778672 0 0x00000000 3211266 jpetazzo 600 393216 2 dest
root@ea319b8ac416:/# ipcs------ Shared Memory Segments --------key shmid owner perms bytes nattch status------ Semaphore Arrays --------key semid owner perms nsems ------ Message Queues --------key msqid owner perms used-bytes messages
user namespace
● no « demo » for this one... Yet!● UID 0→1999 in container C1 is mapped to
UID 10000→11999 in host;UID 0→1999 in container C2 is mapped toUID 12000→13999 in host; etc.
● required lots of VFS and FS patches (esp. XFS)
● what will happen with copy-on-write?– double translation at VFS?
– single root UID on read-only FS?
How does it work?Isolation with cgroups
● memory● cpu● blkio● devices
memory cgroup
● keeps track pages used by each group:– file (read/write/mmap from block devices; swap)
– anonymous (stack, heap, anonymous mmap)
– active (recently accessed)
– inactive (candidate for eviction)
● each page is « charged » to a group● pages can be shared (e.g. if you use any COW FS)
● Individual (per-cgroup) limits and out-of-memory killer
cpu and cpuset cgroups
● keep track of user/system CPU time● set relative weight per group● pin groups to specific CPU(s)
– Can be used to « reserve » CPUs for some apps
– This is also relevant for big NUMA systems
blkio cgroups
● keep track IOs for each block device– read vs write; sync vs async
● set relative weights● set throttle (limits) for each block device
– read vs write; bytes/sec vs operations/sec
Note: earlier versions (pre-3.8) didn't account async correctly. 3.8 is better, but use 3.10 for best results.
devices cgroups
● controls read/write/mknod permissions● typically:
– allow: /dev/{tty,zero,random,null}...
– deny: everything else
– maybe: /dev/net/tun, /dev/fuse
If you're serious about security,you also need…
● capabilities– okay: cap_ipc_lock, cap_lease, cap_mknod, cap_net_admin,
cap_net_bind_service, cap_net_raw
– troublesome: cap_sys_admin (mount!)
● think twice before granting root● grsec is nice● seccomp (very specific use cases); seccomp-bpf● beware of full-scale kernel exploits!
Efficiency
Efficiency: almost no overhead
● processes are isolated, but run straight on the host● CPU performance
= native performance● memory performance
= a few % shaved off for (optional) accounting● network performance
= small overhead; can be optimized to zero overhead
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
Efficiency: storage-friendly
● unioning filesystems(AUFS, overlayfs)
● snapshotting filesystems(BTRFS, ZFS)
● copy-on-write(thin snapshots with LVM or device-mapper)
This is now being integrated with low-level LXC tools as well!
Efficiency: storage-friendly
● provisioning now takes a few milliseconds● … and a few kilobytes● creating a new base image (from a running container) takes a
few seconds (or even less)
Docker
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
What can Docker do?
● Open Source engine to commoditize LXC● using copy-on-write for quick provisioning● allowing to create and share images● standard format for containers
(stack of layers; 1 layer = tarball+metadata)● standard, reproducible way to easily build trusted images
(Dockerfile)
Docker: authoring images
● you can author « images »– either with « run+commit » cycles, taking snapshots
– or with a Dockerfile (=source code for a container)
– both ways, it's ridiculously easy
● you can run them– anywhere
– multiple times
Dockerfile example
FROM ubuntu
RUN apt-get -y updateRUN apt-get install -y g++RUN apt-get install -y erlang-dev erlang-manpages erlang-base-hipe ...RUN apt-get install -y libmozjs185-dev libicu-dev libtool ...RUN apt-get install -y make wget
RUN wget http://.../apache-couchdb-1.3.1.tar.gz | tar -C /tmp -zxf-RUN cd /tmp/apache-couchdb-* && ./configure && make install
RUN printf "[httpd]\nport = 8101\nbind_address = 0.0.0.0" > /usr/local/etc/couchdb/local.d/docker.ini
EXPOSE 8101CMD ["/usr/local/bin/couchdb"]
Yes, but...
● « I don't need Docker; I can do all that stuff with LXC tools, rsync, some scripts! »
● correct on all accounts;but it's also true for apt, dpkg, rpm, yum, etc.
● the whole point is to commoditize,i.e. make it ridiculously easy to use
Containers before Docker
Containers after Docker
What this really means…
● instead of writing « very small shell scripts » tomanage containers, write them to do the rest:– continuous deployment/integration/testing
– orchestration
● = use Docker as a building block● re-use other people images (yay ecosystem!)
Docker: sharing images
● you can push/pull images to/from a registry(public or private)
● you can search images through a public index● dotCloud maintains a collection of base images
(Ubuntu, Fedora...)● satisfaction guaranteed or your money back
Docker: not sharing images
● private registry– for proprietary code
– or security credentials
– or fast local access
● the private registry is availableas an image on the public registry(yes, that makes sense)
Typical workflow
● code in local environment(« dockerized » or not)
● each push to the git repo triggers a hook● the hook tells a build server to clone the code and run « docker
build » (using the Dockerfile)● the containers are tested (nosetests, Jenkins...),
and if the tests pass, pushed to the registry● production servers pull the containers and run them● for network services, load balancers are updated
Hybrid clouds
● Docker is part of OpenStack « Havana »,as a Nova driver + Glance translator
● typical workflow:– code on local environment
– push container to Glance-backed registry
– run and manage containers using OpenStack APIs
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
What's Docker exactly?
● rewrite of dotCloud internal container engine– original version: Python, tied to dotCloud's internal stuff
– released version: Go, legacy-free
● the Docker daemon runs in the background– manages containers, images, and builds
– HTTP API (over UNIX or TCP socket)
– embedded CLI talking to the API
● Open Source (GitHub public repository + issue tracking)● user and dev mailing lists
Docker: the community
● Docker: >160 >170 contributors● latest milestone (0.6): 40 contributors● GitHub repository: >600 >680 forks
Outline
● Why Linux Containers?● What are Linux Containers exactly?● What do we need on top of LXC?● Why Docker?● What is Docker exactly?● Where is it going?
Docker roadmap
● Today: Docker 0.6– LXC
– AUFS
● Tomorrow: Docker 0.7– LXC
– device-mapper thin snapshots (target: RHEL)
● The day after: Docker 1.0– LXC, libvirt, qemu, KVM, OpenVZ, chroot…
– multiple storage back-ends
– plugins
Docker: the ecosystem
● Cocaine (PAAS; has Docker plugin)● CoreOS (full distro based on Docker)● Deis (PAAS; available)● Dokku (mini-Heroku in 100 lines of bash)● Flynn (PAAS; in development)● Maestro (orchestration from a simple YAML file)● OpenStack integration (in Havana, Nova has a Docker driver)● Shipper (fabric-like orchestration)
And many more
Cocaine integration
● what's Cocaine?– Open Source PaaS from Yandex
– modular: can switch logging, storage, etc. without changing apps
– infrastructure abstraction layer + service discovery
– monitoring: metrics collection; load balancing
● why Docker?– Cocaine initially used cgroups
– wanted to add LXC for better isolation and resource control
– heard about Docker at the right time
– uses custom distributed storage instead of Docker registry
●
device-mapper thin snapshots(aka « thinp »)
● start with a 10 GB empty ext4 filesystem– snapshot: that's the root of everything
● base image:– clone the original snapshot
– untar image on the clone
– re-snapshot; that's your image
● create container from image:– clone the image snapshot
– run; repeat cycle as many times as needed
AUFS vs THINP
AUFS● easy to see changes● small change =
copy whole file● ~42 layers● patched kernel
(Debian, Ubuntu OK)● efficient caching● no quotas
THINP● must diff manually● small change =
copy 1 block (100k-1M)
● unlimited layers● stock kernel (>3.2)
(RHEL 2.6.32 OK)● duplicated pages● FS size acts as quota
Misconceptions about THINP
● « performance degradation »no; that was with « old » LVM snapshots
● « can't handle 1000s of volumes »that's LVM; Docker uses devmapper directly
● « if snapshot volume is out of space,it breaks and you lose everything »that's « old » LVM snapshots; thinp halts I/O
● « if still use disk space after 'rm -rf' »no, thanks to 'discard passdown'
Other features in 0.7
● links– linked containers can discover each other
– environment variable injection
– allows to expose remote services thru containers(implements the ambassador pattern)
– side-effect: container naming
● host integration– we ♥ systemd
0.8 and beyond
● beam– introspection API
– based on Redis protocol(i.e. all Redis clients work)
– works well for synchronous req/rep and streams
– reimplementation of Redis core in Go
– think of it as « live environment variables »,that you can watch/subscribe to
● and much more
Thank you! Questions?
http://docker.io/
https://github.com/dotcloud/docker
@docker
@jpetazzo
top related