BMC: Bare Metal Container ＠Open Source Summit Japan 2017

Bare Metal Container
Open Source Summit Japan’17
2017/June/1
1
National Institute of Advanced Industrial
Science and Technology(AIST)
Kuniyasu Suzaki

Contents
• Background of BMC
– Drawbacks of container, general kernel, and accounting.
• What is BMC?
• Current implementation
• Evaluation
• Extension
– NVIDIA Docker, Moby, Intel Clear Container, etc.
• Conclusions
2

Background of BMC 1/3
Drawback of Container
• Container technology (Docker) becomes popular.
– Docker offers an environment to customize an application easily.
– It looks like to be good for an application, but it is a server centric.
• It does not allow to change the kernel.
– Kernel options passed through /sys are not effective.
• Some applications cannot run on Docker.
– DPDK on Docker does not work on some machines, because it
depends on “igb_uio” and “rte_kni” kernel modules.
• Some provider offers the kernel which can treat DPDK on Docker,
but it is case by case solution. It is not fundamental solution.
3

4
Container is a jail for a kernel optimizer.

5
Container is a jail for a kernel optimizer.
HPC users want to optimize the kernel for their
applications. Kernel is a servant.
Container way is not fit for them.

General kernel leads weak performance
• Arrakis[OSDI’14] showed that nearly 70% of network latency
was spent in the network stack in a Linux kernel.
• Many DB applications (e.g., Oracle, MongoDB) reduce the
performance by THP (Transparent Huge Pages) which is
enabled on most Linux distributions.
6

General kernel leads weak performance
• Arrakis[OSDI’14] showed that nearly 70% of network latency
was spent in the network stack in a Linux kernel.
• Many DB applications (e.g., Oracle, MongoDB) reduce the
performance by THP (Transparent Huge Pages) which is
enabled on most Linux distributions.
7
It is not fundamental solution.
HPC users want to optimize the kernel for their
applications. Kernel is a servant.

Power consumption for each application
• Current power measurement is coarse.
– Power Usage Effectiveness: PUE only shows usage of data-center
scale.
– Current power consumption is theme for vender and administrators
• Users have no incentive for low power, even if they make a
low power application.
– Current accounting is based on time consumption.
8

Power consumption for each application
• Current power measurement is coarse.
– Power Usage Effectiveness: PUE only shows usage of data-center
scale.
– Current power consumption is theme for vender and administrators
• Users have no incentive for low power, even if they make a
low power application.
– Current accounting is based on time consumption.
9
There is no good method to measure power
consumption “for an application”.
No accounting which considers power consumption.

What is BMC?
• BMC(Bare-Metal Container) runs a container (Docker) image
with a suitable Linux kernel on a remote physical machine.
– Application on Container can change kernel settings and machine
which fit for it. BMC extracts the full performance.
– On BMC, the power on the machine is almost used for the application.
• BMC tells the power usage on each machine architecture. Users can
know which architecture is good for their application.
10
BMC offers incentives for an application to
customize kernel and select machine architecture

What is BMC?
• BMC(Bare-Metal Container) runs a container (Docker) image
with a suitable Linux kernel on a remote physical machine.
– Application on Container can change kernel settings and machine
which fit for it. BMC extracts the full performance.
– On BMC, the power on the machine is almost used for the application.
• BMC tells the power usage on each machine architecture. Users can
know which architecture is good for their application.
11
BMC offers incentives for an application to
customize kernel and select machine architecture

machine
kernel
container manager
Server Centric Architecture
Traditional Style
(Ex: container)
Invoke app.
Power always up
Admin’s
Space
User’s
Space
app
container
app
container
app
container
Comparison
Pros:
• Multi Tenant (high CPU utilization)
Cons:
• Kernel is not replaced.
• No information of power consumption
Pros:
• Apps select a kernel & hardware (low power to high spec).
• Apps occupy the machine and extract the performance.
Cons:
• Set up overhead (Rebooting)
Boot the kernel & app.
BMC
machine machine machine
kernel
app
container
kernel kernel
Application Centric Architecture
Select a kernel
Select a physical
machine
BMC manager
Remote Machine
management
(WOL, AMT, IPMI)
network
bootloader
network
bootloader
network
bootloader
Power frequently up/down
app
container
app
container

Node-1
Docker
Hub
BMC
Hub
BMC Manager
client
BMC Command
#bmc run “Docker-img” “kernel” “initrd” “app and argument”
HTTPS (apache)
iPXE script
kernel & initrd
kernel & initrd
IP address
(bmc-ID)
NFS mount or download to RAM FS
docker image
Docker Image
ssh
ssh pub-key
cloud-init
+ bmc tools (heatbeat)
+ sshd
+ ssh pub-key
iPXE
Power On (WOL, AMT, IPMI)
Download iPXE script
Download kernel & initrd
NFS mount or download to RAM FS
request ssh connection
①
②③
④
⑤
Power Off (shutdown command, AMT, IPMI)
⑥
⑦
⑧
iPXE
Kernel & initrd
(IP3)
(MAC or IP1)
(IP2)
(Linux or IP1)
Procedure to execute BMC command

Remote Machine Boot Procedure
1. Power-on a node machine with Remote Machine
Management (WOL, Intel AMT, IPMI)
2. Network Boot Loader (iPXE)
– Get kernel and intird from a HTTP/HTTPS server.
3. The downloaded initrd mounts a Docker image.
• NFS mode
• RAM FS mode
4. Boot procedure in a Docker image
– Fortunately, Docker image keeps boot procedure.
5. SSH is connected from BMC command
– Run an application.
14

Remote Machine Management
WOL Intel AMT IPMI
Protocol
Magic Packet
(MAC address)
HTTPS
(IP address)
RMPC
(IP address)
Power-On ✔ ✔ ✔
Power-Off × ✔ ✔
Security × Password Password
Comment
Most PCs
have WOL.
High level Intel
machine
Server Machine
(Slow BIOS)
15

Network Boot Loader
• PXE is the most famous, but it is limited for LAN, because
it depends on “magic packet” of Layer 2.
• BMC uses iPXE which download “kernel” and “initrd” from
HTTP/HTTPS.
• The iPXE downloads kernel and initrd.
16
#!ipxe
ifopen net0
set net0/ip 192.168.0.101
set net0/netmask 255.255.255.0
set net0/gateway 192.168.0.1
set dns 192.168.0.1
:loop
chain http://192.168.0.200/cgi-bin/baremetal.ipxe || goto waiting
exit
:waiting
sleep 1
goto loop
– iPXE is custimzed by its
scripting language. BMC
uses it.

How to boot OS (Linux)
• The downloaded “initrd” is customized to mount an Docker
image. It offers 2 mount methods.
– NFS mode
• Download necessary data only and fast boot, but it needs to download data to
run applications after boot.
– RAMFS mode
• Download full Docker image and slow boot, but application runs fast after boot.
• Boot procedure in the Docker image.
– An Docker image keeps boot procedure for each application because each
application package designed to include them.
– BMC utilizes these boot procedures to rum daemons, such as the SSH,
because an application in the Docker image is executed by remote
procedure calls from BMC manager.
17

Power Consumption
• Each node has power meter “WattChecker”.
• WattChecker measures power consumption from the
power-on caused by WOL,AMT, or IPMI.
• BMC manager keeps the log of the power
consumption.
• The log is used for power accounting.
• It is coarse, but it shows affinity between application
and architecture.
18

Current Implementation
• Current BMC Manager is implemented with shell script.
– 4, 500 LOC.
• Power consumed on each node is measured by
WattChecker.
• We have tried several machines as BMC nodes.
– From Atom to Xeon.
– Application can select machine considering power
consumption.
19

Spec of Test Machines
Remote
machine
manage
ment
CPU,Core/thread,
Clock (Burst time),
Power
Logical
performance
GFLOPS
(Burst time)
Issue
date
Memory NIC
(queue)
Low Power
Intel NUC 5CPYH
WOL Celeron (N3050),2/2, 1.6
(2.16)GHz,8W
6.4
(8.6)
2015 8GB RealTek r8169
(1)
NotePC
Lenovo ThinkPAD
T430s
Intel AMT i7 (3520M)2/4,
2.9(3.6)GHz, 35W
46.4
(57.6)
2012 16GB
Intel e1000
(1)
DesktopPC
Dell Optiplex 960
Intel AMT Core 2Quad (Q9400)
4 /4, 2.66GHz,95W
42.656 2008 16GB
Intel e1000
(1)
Server
Dell PowerEdge
T410
IPMI Xeon (X5650)
6/12,2.66(3.06)GHz,95W
63.984
(73.44)
2010 8GB
Broadcom
NeXtreme II(8)
20

Boot performance (overhead)
21
Network PowerTime
NFSRamFS
• They are BMC’s overhead.
• The performance improved by optimization must surpass the overhead.
Time(s) Power (j) Traffic (MB)
NFS mode
Celeron 35.4 242 49.5
Core2
Quad
28.1 1,773 49.3
I7 20.9 481 49.1
Xeon 92.6 9,932 49.8
RAMFS mode
Celeron 55.6 402 92.9
Core2
Quad
40.0 2,493 92.8
i7 34.3 775 92.8
Xeon 102.7 11,015 92.6
46 times
oo

Tested Application and Optimization
• This presentation shows the result of Matrix
multiplication with/without Hyper Threading.
– The experiment measured the time for 10 times of matrix
multiplications on OpenBlas optimized for each machine.
22
Application Optimization
Matrix Multiplication with
OpenBlas
Hyper Threading off
Redis benchmark Transparent Huge Pages off
Apache benchmark Receive Flow Steering off

Performance Difference
10 times of matrix multiplications [12800:12800]
on OpenBlas optimized for each machine. .
23
() shows the rate from logical performance
• The results show no hyper threading is better.
• Xeon shows the best performance, but i7 shows cost effective.
Time (s) Power (j) GFLOPS Power/(GFLOPS*time)
Celeron 12,783.8 125,084
2.99
(34.7%)
3.27
Core2Quad 1,060.2 140,346
39.8
(93.3%)
3.32
i7
HTT-on
961.4 55,315
43.8
(76.0%)
1.31
i7
HTT-off
827.1 45,364
50.9
(88.4%)
1.08
Xeon
HTT-on
945.6 211,908
44.6
(60.7%)
5.02
Xeon
HTT-off
698.9 151,760
60.5
(82.4%)
3.59
Celeron
(Atom Core)
is not cost
effective
CPU.

Performance improvement which compensates
the boot overhead
Boot
overhead
Improvement
at [6400:6400]
Improvement
at [12800:12800]
Time (sec)
i7 35.4 15.9 134.3
Xeon 108.0 29.8 246.7
Power (joule)
i7 1,805.3 1,150 9,951
Xeon 11,274.5 6,792 60,148
24
• Overheads caused by booting were compensated before
[12800:12800].

BMC Extension
• NVIDIA-Docker
• Moby
• Intel Clear Containers
• OSes for Container
25

Extension for “NVIDIA-Docker”
• NVIDIA-Docker runs CUDA applications.
– It manages CUDA and driver of NVIDIA GPU.
• A suitable CUDA is added Docker image
automatically.
• So, users do not need to install CUDA.
– Users don’t need to care about CUDA version.
• BMC is now customizing to add CUDA
which matches to NVIDIA-driver.
– The target is TensorFlow on NVIDIA-Docker.
• https://hub.docker.com/r/tensorflow/
26

Moby
• Moby is a framework to assemble container systems.
• Moby allows to run “redis” container without Docker!
– Why do we need “containerd” and “LinuxKit”?
– Redis should run on native Linux kernel?
27
https://www.slideshare.net/Docker/dockercon-2017-general-session-day-1-solomon-hykes-75362520

Moby
• Moby is a framework to assemble container systems.
• Moby allows to run “redis” container without Dock!
– Why do we need “containerd” and “LinuxKit”?
– Redis should run on native Linux kernel?
28
https://www.slideshare.net/Docker/dockercon-2017-general-session-day-1-solomon-hykes-75362520
iPXE
Normal Linux
“RedisOS”
on normal Linux
BMC

Intel Clear Container
• Intel Clear Container is a counter-thesis of container. It
encourages to use virtualization (Intel VT).
– Intel insists that Intel VT and Linux’s boot (< 200ms!) are fast.
Why don’t you use virtualization? Virtualization offers stronger
isolation.
• BMC’ s offers counter-thesis against Intel Clear Container.
– If Linux’s boot is fast, why don’t you use native Linux?
– BMC offers more flexibility of kernel customization.
29

Current BMC target
• Many Docker OSes are proposed.
–CoreOS
–RancherOS
–Snappy Ubuntu Core
–RedHat Project Atomic
–Mesosphere DC/OS
–VMware Photon
• BMC can use these Linux kernels.
–BMC only requires kernel and container image.
30

Related works
• Triton [Joyent’s product]
– Triton = Docker + SmartOS.
• In order to optimize, user needs to customize SmarOS.
• LinuxBIOS/BProc Cluster[HPCS’02]
– Testbed for kernel test. It is not so easy to implement because it
requires to replace BIOS.
• SLURM[ICDCN’14]
– Measure power consumption for an application. It depends on
function to measure power (Intel RAPL: Running Average Power
Limit, or CRAY machine).
31

Conclusions
• BMC (Bare-Metal Container) runs a container
(Docker) image with a suitable Linux kernel
on a remote physical machine.
• The overhead of BMC was compensated by
the improved performance of applications.
• Official HP: http://www.itri.aist.go.jp/cpc/research/bmc/
• Docker Image for BMC manager:
https://hub.docker.com/r/baremetalcontainer/
• Source Code: https://github.com/baremetalcontainer
32

BMC: Bare Metal Container ＠Open Source Summit Japan 2017

More Related Content

What's hot

Similar to BMC: Bare Metal Container ＠Open Source Summit Japan 2017

More from Kuniyasu Suzaki

Recently uploaded

BMC: Bare Metal Container ＠Open Source Summit Japan 2017