ai, app-defined os, and the ossified kernel
TRANSCRIPT
![Page 1: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/1.jpg)
AI, app-defined OS, & the ossified Linux kernel
Felix Xiaozhu Lin 林小竹
1
![Page 2: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/2.jpg)
Summary
Describe recent trends of OS research
Present three cases of our investigation
Share personal reflection on OS research
2
![Page 3: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/3.jpg)
Self intro• 2014 – now. Asst. prof, Purdue ECE
• 2014 PhD in CS. Rice U (advisor: Lin Zhong)• Thesis: OS for mobile computing
• 2008 MS + BS. Tsinghua U
3
XSEL & Friends, 2017
Crossroads Systems Exploration Lab
![Page 4: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/4.jpg)
AI + big data: the workloads in our era
How does OS respond?4
![Page 5: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/5.jpg)
OS as in textbooks
• Resource management
• Abstraction
• Protection
• Multiplexing
5
![Page 6: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/6.jpg)
6
SOSP 1995
Hardware
NN Stream M/R
Mgmt Abstraction
Protect Multiplex
OS in 2018
Hardware
Mgmt Abstraction
Protect Multiplex
OS in 2000
![Page 7: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/7.jpg)
Two premises of our research
1. Important apps define OSes
2. The Linux kernel is the new firmware
7
![Page 8: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/8.jpg)
Like it or not …
• Following the model: widely adopted• Various ML frameworks• Android• DPDK• RDMA
• Against the mode: less adopted• Library OS• Multikernel• Unikernels• (… and all kinds of research proposals)
8
![Page 9: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/9.jpg)
This talk: our exploration in three cases
• App-defined OSes• Case 1: Memory mgmt
• Case 2: Storage
• The ossified Linux Kernel• Case 3: Transkernel
9
![Page 10: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/10.jpg)
Case 1: App-defined memory management
Exploiting hybrid memory for stream analytics
10
![Page 11: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/11.jpg)
High input throughput: millions of events per sec
Low delay processing: sub-seconds
IoT Data centers Humans
11
Background: Stream analytics in 1 minute
![Page 12: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/12.jpg)
Background: streaming pipeline
IngressGroup by
wordCount word occurrences
12
Operators Computations consuming & producing streams
Pipeline A graph of operators
WordCount
Window egress
![Page 13: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/13.jpg)
Background: streaming pipeline
Ingress
Group by word
Count word occurrences
13
Operators Computations consuming & producing streams
Pipeline A graph of operators
TopKwords
Window
SentimentScoring Join
Sentiment Detection
![Page 14: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/14.jpg)
14
Hybrid high-bandwidth memory
![Page 15: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/15.jpg)
Hybrid high-bandwidth memory
• Tradeoffs: capacity vs bandwidth
• Untraditional memory hierarchy
– No latency benefit (Unlike SRAM+DRAM)
– Configurable: sw-managed or hw-managed15
CPU High-bw mem (HBM)
DRAM
16 GB375 GB/s
96 GB80 GB/s
![Page 16: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/16.jpg)
Accelerate stream analytics with HBM?• The most expensive stream operators: grouping
16
Grouping in Hash High bandwidth memory
High capacity demand Capacity limited
Extensive random access ❤ Seq access + parallelism
IngressGroup by
wordCount word occurrences
TopK wordsWindow
Grouping traditionally built as HashMismatch HBM!
![Page 17: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/17.jpg)
HBM can accelerate grouping!
• … with an unconventional algorithm choice
• Grouping: hash → sort
• Sort is worse than Hash on algorithmic complexity• O(NlogN) vs. O(N)
• Yet, Sort beats Hash with …• Abundant mem bw• High task parallelism• Wide SIMD (avx512)
17
![Page 18: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/18.jpg)
A bold design: only use HBM for in-mem index of parallel sorting
18
Mem mgmt
HBMDRAM
Streamingdata
In-memindex
DataBundles
![Page 19: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/19.jpg)
Memory allocator: balancing two limited resources• Prevent either from becoming the bottleneck
• A single knob: where to allocate new index: HBM or DRAM?
19DRAM Bandwidth
HB
M C
apac
ity
![Page 20: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/20.jpg)
Memory allocator: balancing two limited resources• Prevent either from becoming the bottleneck
• A single knob: where to allocate new index: HBM or DRAM?
20
Low-demandfor both
High-demandfor both
DRAM Bandwidth
HB
M C
apac
ity
![Page 21: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/21.jpg)
Memory allocator: balancing two limited resources• Prevent either from becoming the bottleneck
• A single knob: where to allocate new index: HBM or DRAM?
21
Low-demandfor both
High-demandfor both
DRAM Bandwidth
HB
M C
apac
ity
![Page 22: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/22.jpg)
Memory allocator: balancing two limited resources• Prevent either from becoming the bottleneck
• A single knob: where to allocate new index: HBM or DRAM?
22
Low-demandfor both
High-demandfor both
DRAM Bandwidth
HB
M C
apac
ity
![Page 23: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/23.jpg)
Case 1: First streaming engine optimized for hybrid memory
• Works on real hardware
• Best stream analytics performance on a single machine
• Generic memory allocators are inadequate!
25
![Page 24: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/24.jpg)
This talk: our exploration in three cases
• App-defined OSes• Case 1: Memory mgmt
• Case 2: Storage
• The ossified Linux Kernel• Case 3: Transkernel
26
![Page 25: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/25.jpg)
Case 2: App-defined storage
A data store for video analytics
27
![Page 26: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/26.jpg)
Pervasive cameras. Large videos.
• 130M surveillance cameras shipped per year
• Many campuses run > 200 cameras 24x7
• A single camera produces 24 GB video per day
28
![Page 27: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/27.jpg)
Pervasive cameras. Large videos.
• 130M surveillance cameras shipped per year
• Many institutions run > 200 cameras 24x7
• A single camera produces 24 GB video per day
29
![Page 28: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/28.jpg)
Retrospective video analytics in 30 seconds• A query: “find all white buses appeared yesterday”
• A cascade of operators
• Query picks operator accuracies• Lower accuracy → lower cost → faster execution
30Image credits: NoScope: Optimizing Neural Network Queriesover Video at Scale, Daniel Kang, John Emmons, Firas Abuzaid, Peter Bailis, Matei Zaharia, VLDB 2017
Frame diff detector
Specializedneural net
Fullneural net
~10,000x ~1,000x ~10x
![Page 29: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/29.jpg)
Need for a video store for retrospective analytics
31
Ingestion Storage Retrieval Consume
VideoData
Operator@ accuracy
Query
“Aren’t there many video databases already?”Yes. But they are for human consumers.Not for algorithmic consumers.
![Page 30: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/30.jpg)
Central design issue: choose storage formats catering to operators
• Video format: resolution, frame rate, crop, compression level, color channel… Many knobs!
32
![Page 31: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/31.jpg)
Central design issue: choose storage formats catering to operators
• Video format: resolution, frame rate, crop, compression level, color channel… Many knobs!
• Higher accuracy → richer format & higher cost
33Actual tuning of formats is complex!
200p 1FPS(~100KB/s)
720p 30FPS(~400KB/s)
![Page 32: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/32.jpg)
Key problem:What video formats to store?
1. Retrieving formats must be fast enough to feed operators
2. Formats must be rich enough to satisfy all accuracy needs
34
![Page 33: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/33.jpg)
35
Storageformats
<motion,0.95>
<motion, 0.7>
…
<OCR, 0.95>
<OCR, 0.90>
…
<NN, 0.95>
<NN, 0.80>
Video consumers<op, accuracy>
Video data path
Retrieving & decoding slows down operators
Storing one unified format?
![Page 34: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/34.jpg)
Storing all needed formats?
36
Storageformats
<motion,0.95>
<motion, 0.7>
…
<OCR, 0.95>
<OCR, 0.90>
…
<NN, 0.95>
<NN, 0.80>
Video consumers<op, accuracy>
Video data path
Ingestion & storage are expensive
![Page 35: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/35.jpg)
Proposal: Store minimum necessary formats
37
Consumption formats
Storageformats
<motion,0.95>
<motion, 0.7>
…
<OCR, 0.95>
<OCR, 0.90>
…
<NN, 0.95>
<NN, 0.80>
Video consumers<op, accuracy>
Video data path
![Page 36: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/36.jpg)
Proposal: Store minimum necessary formats
38
Consumption formats
Storageformats
<motion,0.95>
<motion, 0.7>
…
<OCR, 0.95>
<OCR, 0.90>
…
<NN, 0.95>
<NN, 0.80>
Video consumers<op, accuracy>
Video data path
Key Challenge: Too many knobs to tune!
![Page 37: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/37.jpg)
Proposal: Store minimum necessary formats & derive them automatically!
39
<motion,0.95>
<motion, 0.7>
…
<OCR, 0.95>
<OCR, 0.90>
…
<NN, 0.95>
<NN, 0.80>
Backward derivation of configuration
Video data path
Consumption formats
Storageformats
Video consumers<op, accuracy>
![Page 38: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/38.jpg)
TargetAccuracy
1x
10x
100x
1000x
1
0.9
5
0.9
0.8
Qu
ery
sp
ee
d (
x re
alti
me
)Benefit: faster analytics
40
![Page 39: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/39.jpg)
TargetAccuracy
1x
10x
100x
1000x
1
0.9
5
0.9
0.8
Qu
ery
sp
ee
d (
x re
alti
me
)
OursOne unified format
Benefit: faster analytics
Why?• Lower accuracy → analytics tolerates cheaper video formats • Video retrieval should be proportionally cheaper!
41
![Page 40: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/40.jpg)
jackson
1x
10x
100x
1000x
1
0.9
5
0.9
0.8
miami
1
0.9
5
0.9
0.8
tucson
1
0.9
5
0.9
0.8
dashcam
1
0.9
5
0.9
0.8
park
1
0.9
5
0.9
0.8
airport
1
0.9
5
0.9
0.8
OursStore one format
Qu
ery
sp
ee
d (
x re
alti
me
)Benefit: faster analytics
16x
Why?• Lower accuracy → analytics tolerates cheaper video formats • Video retrieval should be proportionally cheaper!
42
![Page 41: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/41.jpg)
Case 2: First data store for retro. video analytics• Baking analytics demands into storage decisions
• Showing clear advantage over generic file systems or storage layers
45
![Page 42: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/42.jpg)
This talk: our exploration in three cases
• App-defined OSes• Case 1: Memory mgmt
• Case 2: Storage
• The ossified Linux Kernel• Case 3: Transkernel
46
![Page 43: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/43.jpg)
Case 3: enhancing the ossified Linux kernel
A new OS structure for device driver offloading
48
![Page 44: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/44.jpg)
A commodity kernel still irreplaceable
• The major reason: device drivers• Linux: a common environment for all drivers
• >70% Linux kernel source are device drivers
• All other OS functions have alternatives!
491. Understanding Modern Device Drivers, Asim Kadav and Michael Swift. ASPLOS’12.
![Page 45: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/45.jpg)
• Mobile/IoT run frequent, ephemeral tasks• Android smartwatch: each minute
• Each task is short-lived• Smartwatch: 10 secs
• Background task: < 1 sec
50
Problem: suspend/resume in ephemeral tasks
![Page 46: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/46.jpg)
Under the hood…
51
Time
Device
Suspend
Wakeup
User Task
Sleep
Device
Resume
Thaw user
Freeze user&fs
![Page 47: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/47.jpg)
Suspend/Resume Is Slow
52
Nexus 5 Gear Note 4 Panda
Suspend 119 ms 191 ms 231 ms 262 ms
Resume 88 ms 159 ms 316 ms 492 ms
Why? IO devices are slow
![Page 48: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/48.jpg)
Proposal: run these driver paths on a low-power tiny core
54
CPU
2.5GHz 50MHz
![Page 49: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/49.jpg)
What are these tiny cores?
55
![Page 50: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/50.jpg)
What are these tiny cores?
56
![Page 51: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/51.jpg)
What are these tiny cores?
57
![Page 52: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/52.jpg)
How can a tiny core help?
• Idling more efficiently• 3 mW vs 30 mW
• Executing kernel more efficiently• Small code working set
• Less predictable control flow
58
1. J. Mogul, J. Mudigonda, N. Binkert, P. Ranganathan, and V. Talwar. Using asymmetric single-isacmps to save energy on operating systems. Micro, IEEE, 2008
![Page 53: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/53.jpg)
The workflows
59
Time
Device
Suspend
Wakeup
CPU
User Task
Sleep
Device
Resume
Thaw user
Freeze user&fs
Current
![Page 54: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/54.jpg)
The workflows
60
Time
Device
Suspend
Wakeup
CPU
User Task
Sleep
Device
Resume
Thaw user
Freeze user&fs
CPU Tiny Core
Current Proposed
![Page 55: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/55.jpg)
How?
• The tiny core is very wimpy• Essentially a microcontroller
• Has a different ISA• More aggressive than big.LITTLE
• Re-engineering the kernel?
• “Linux kernel is the new firmware”
61
Device
Drivers
Linux Kernel
DRAM IO
CPU
![Page 56: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/56.jpg)
Proposal: use dynamic binary translation• Empower the tiny core to execute unmodified
kernel binaries
62
Dynamic
Binary
Translation
Device
Drivers
Linux Kernel
DRAM IO
CPU
![Page 57: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/57.jpg)
Dynamic binary translation in 30 sec
• The technology behind QEMU and a PS emulator
63
![Page 58: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/58.jpg)
“This sounds impractical”
• Dynamic bin translation (DBT) is known expensive• > 10x overhead
• No one runs DBT on microcontrollers• Always weak over strong
65
![Page 59: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/59.jpg)
Solution: the Transkernel model
66
Key ideas:
1. Translate most of driver code
2. Emulate the remaining kernel infrastructure
Essentially a tiny VM for drivers!
CPU Tiny core
Device
Drivers
Linux
kernel
DRAM
Emu
Dynamic
Binary
Translation
IO
Translated
code
A transkernel
![Page 60: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/60.jpg)
Is this practical?
67
![Page 61: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/61.jpg)
Only through careful optimization!
68
![Page 62: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/62.jpg)
Case 3: Transkernel: tiny VMs for device driver paths
First DBT engine running on a microcontroller-like core!
Demonstrated:
• One can use DBT for efficiency gain
• The ossified kernel can be tamed!
69
![Page 63: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/63.jpg)
This talk: our exploration in three cases
• App-defined OSes• Case 1: Memory mgmt for stream analytics
• Case 2: Storage for retro. video analytics
• The ossified Linux Kernel• Case 3: Transkernel
70
![Page 64: AI, app-defined OS, and the ossified kernel](https://reader030.vdocuments.net/reader030/viewer/2022032721/623d7e4b30633a2a153f7028/html5/thumbnails/64.jpg)
Reflection: a tale of two inquiries
• App-defined OS• Brave new world
• Usefulness depends on the target apps
• OS as a tool: serving AI and big data
• Kernel as the new firmware• Still lots of opportunities
• Unique constraints. Require unconventional techniques
• Like fixing an engine of an airplane in the air
• OS as a craft. Fun!
72