240705-랩소개-전산학부-석사신입생.pptx

Introduction about ANLAB at KAIST.

1. ANLAB 연구실 소개 2024-07-05

2. Professor: 문수복 (Sue Moon) 2

3. 장거리 통신망 << 데이터센터 내부망 다우 존스 인덱스 변화 AT&T: 1916 년 ~ 2004 년 / 2005 년 ~2015 년 Verizon: 2004 년 ~ 현재 Microsoft: 1999 년 ~ 현재

4. 전력소모로 예상되는 데이터센터 성장세 2018.9.12. Nature

5. Networking System Technologies for Software-Defined Infrastructure MAGA companies build datacenters based on proprietary technologies MS Azure: FPGA-baesd SmartNIC Amazon: AWS Google: Search Engine & YouTube Apple: mostly for iCloud support Datacenters operate at too large a scale: MS Azure 100K different backbone routes 400K interfaces 6K independent SRLGs (Shared Risk Link Group) Challenges Too many factors to track Need flexible environment to iterate 5 Quotes from Dave Maltz’s talk at MSR Faculty Summit 2021

6. Ongoing Projects / Interests Keunhong Lee Asynchronous NFV framework (using Rust) Wonsup Yoon w/ Prof. Youngjin Kwon Disaggregated Data Center (DDC) Jisu Ok RDMA as a new networking primitive Soonhyeon Kwon Polarization in online social network Seongkyu Lee Task scheduling in distributed systems 6

7. 7 Adapting Async/Await for Network Systems Event-driven Paradigm 3 distinct event handlers These handlers are linked by a shared state machine Async/await paradigm 1 asynchronous task The reusable combinator then , represents the procedure of a handshake process https://www.techopedia.com/definition/10339/three-way-handshake Fall, Kevin R., and W. Richard Stevens.  Tcp / ip illustrated . Vol. 1. Addison-Wesley Professional, 2012. onSynRecv onEstablish onConnect start(). . then ( sendSyn ) . then ( recvSynAck ) . then ( sendAck )

8. Memory Disaggregation using LibOS 8 Memory disaggregation replaces landscape of DC Separating compute and memory node Scaling resources independently LibOS benefits from Fast PF handling app-aware opt. Compatibility with existing program DiLOS : LibOS solution (EuroSys’23) Enables high perf. & compatibility Outperforms SotA systems up to 2.2x ( Fastswap ) and 1.5x (AIFM) Towards Microsecond-scale Tail Latency under MD Eliminates HoLB thru asynchronous page fault handling Boosting context switching perf. through user-level scheduling

9. Expanding RDMA beyond Intra-DCN 9 RDMA is now a primary networking primitive for intra-DCN RDMA’s promise: high throughput, low latency, near-zero CPU consumption Intra-DCN networking: RTT is typically 1s ~ 100s microseconds Inter-DCN also requires high performance networking Cloud operators deploy multiple DCs within the same region Inter-DCN networking: path RTT in milliseconds-scale Research questions How can we emulate RDMA networking over the long path delay? How does RDMA behave when the path RTT goes up to ms -scale?

10. Polarization analysis in social media 10 Social media data collection and preprocessing Data collection from multiple media Multimodal data analysis Operationalization of polarization and behavior Transformer Moral Foundation Dictionary(MFD)

11. My Track Record Have covered all top conferences SIGCOMM, NSDI, INFOCOM, WWW, IMC, CoNEXT , … Have served in all top conference TPCs SIGCOMM, NSDI, INFOCOM, WWW, IMC, CoNEXT , … Have won best papers / posters IMC, EuroSys , APNet / SIGCOMM Serving in award committees IEEE Internet Award, SIGCOMM Rising Star Award, COSN Best Paper Award Writing recommendation letters for Top schools, green cards, … My papers have been widely cited & read in classrooms Total citation count = 24,000+, h-index = 45 Remaining challenges H-index = 50 ACM/IEEE fellow Changing SCI-journal-based evaluation system 20 PhDs? 11

12. Alumni Brain Drain: 2/3 Currently Abroad 12 AT&T, Telefonica intern MPI post-doc KAIST CT professor KAIST CS professor IBS Data Science CI MPI Director Telefonica intern Telefonica researcher QCRI researcher SMU Professor Indiana University Prof. UCLA IPAM intern MSR Camb UK intern 국가보안연구소 ETRI 연구원 Intel Camb UK intern SKT 상무 UCSD intern MSR Redmond intern MSR Camb UK post-doc Intel researcher Google SW engineer MPI-SWS professor Rubrik Engineer MSR Camb UK intern Lablup founder Yahoo! Labs intern MPI-SWS intern 삼성전자 Qualcomm UC Berkeley intern UC Berkeley post-doc Nefeli Networks Alkira NVIDIA

13. Current Lab Members 13 admin staff Ph.D Ph.D Ph.D MS Undergrad MS (Starting MS this fall) Undergrad

14. Miscellaneous Info 연구실 모토 “ 적은 돈으로 해결할 수 있는 일은 돈으로 해결하자 ” “ 포스터 그만 쓰자 . Full paper 좀 쓰자 ” “ 큰 그림 그리자 ” “Leave a mark” (circa 2010) 랩 운영 방침 랩장 : 최소한의 랩운영을 위한 coordinator 1 주에 한 번 연구미팅 / 격주 랩미팅 Retreat 일년에 2 번 brainstorming 과제 = 연구 14

15. Lab Track Record +/ Avg time to PhD = 5~6 +/- Never more than 10 students at any time + /- All PhD graduates/candidates did overseas internships +/- Self-motivation most important +/- No set office hours 15

16. Q & A 개별 / 졸업연구를 고민하는 학생들을 위한 간추린 자료 http://sbmoon.tistory.com/245 내가 생각하는 전산학과 대학원 진학 준비 http://sbmoon.tistory.com/231 처음 논문쓰는 학생들을 위한 논문 작성법 http://sbmoon.tistory.com/241 석사논문발표 준비하는 학생들을 위한 가이드 http://sbmoon.tistory.com/243 가을학기 석사 신입생 모집 중 ( 잔여 TO: 1 명 ) If you are interested, feel free to contact: Prof. Moon - [email protected] Student Representative ( 랩장 ) – [email protected]

17. 홈페이지 캘린더 보고 빈 시간에 면담 요청하세요 ! 17 http://an.kaist.ac.kr/~sbmoon/#schedule

18. Thank you! https://an.kaist.ac.kr/ 18

19. BACKUP SLIDES 19

20. 미국 국방부 프로젝트로 시작 Courtesy Arturo Contreras

21. 1982 년 국내 첫 인터넷 연결 구성도 한국인터넷진흥원

22. ANLAB 연구실 소개 첨단네트워킹 연구실 소개

23. ANLAB 연구실 소개 첨단네트워킹 연구실 소개

24. Keunhong: Design Reusable Building Blocks for NFs Challenges: Convert event-driven programs to asynch/await paradigm Signaling method - Interrupt - Polling Data sharing method Pan-packet static storage Per-packet dynamic storage Combinators Demux / Mux Queue Batched computation 27

25. Type System & Networking 28 Rust Programming Language Pros Type-safe (no undefined behavior) Fast (no overhead) Cons Hard to program Slow (when ownership type is not well-defined)

26. Middleboxes or Network Functions 29 Firewall [From Justine Sherry’s SIGCOMM 2012 Talk]

27. Moving Enterprise Network to Cloud 30 [From Justine Sherry’s SIGCOMM 2012 Talk]

28. SA Top lab to focus on research!

29. Type System & Networking 32 Memory-safe Network Services Through A Userspace Networking Switch (Kai Lüke , 2018 alumni) Programmed by Rust language to ensure memory safety http://pothos.blogsport.eu/files/2019/01/msc_thesis_memory-safe_network_services_userspace_switch.pdf

30. Type System & Networking 33 Enveloping Implicit Assumptions of Intrusive Data Structures within Ownership Type System (PLOS 19) https://ess.cs.uni-osnabrueck.de/workshops/plos/2019/preprint/plos19-lee.pdf

31. Type System & Networking 34 Network Flow Processing Using Rust (ongoing) Key Idea Static typing for network flows Full utilization of compiler optimization Dynamic offloading without abstraction cost To CPU (multi-threading) To GPU To FPGA To remote machine https://dl.acm.org/citation.cfm?id=319166

32. Adapting Async/Await for Network Systems 35 Challenges of adopting async/await to network systems Lack of poll-aware scheduler Network systems eagerly use polling to reduce latency, however polling wastes CPU cycles if not properly scheduled Solution Distinguish polling and non-polling task types on separate run Q. A tailored scheduling algorithm Manually embedded high-performance techniques (e.g. zero-copy API, direct memory access, and resettable timer) High-performance techniques are entangled with network systems so that each technique implementation is less reusable. Solution Each protocol layer denotes its required access on packets (e.g. read access, write access, or access to create new references). The framework determines whether copy is required or not. Timer management task is scheduled as a poll-type task. The framework’s scheduler determines when to iterate through pending timers. Unhandled unique data movement patterns of network systems (e.g. batch, demux , mux, broadcast) Network systems have unique data movement patterns but some of them are not covered by existing framework. Solution Dedicated combinator implementations for each data movement pattern

33. UDP Echo server w/ 64B payload Legend SOTA: Linux (with all zero-copy solutions), Demikernel (with DPDK) 1: Our network stack w/ high-performance techniques 2: 1 + Poll-aware scheduler 3: 1 + Dedicated combinators 4: 1 + Both 36 Adapting Async/Await for Network Systems SOTA 1 2 3 4 Program your network stack within the async/await paradigm. Let the framework enhance the performance.

34. Disaggregated Data Center 37

35. Memory Disaggregation using LibOS 38 https://dl.acm.org/doi/abs/10.1145/3476886.3477507 LibOS -based memory disaggregation enables Fast page fault handling Application-aware memory prefetching Compatibility with existing program binaries

36. Online News and Opinion Dynamics 39 북핵실험 아이스하키 단일팀 경선 / 대선 연예계 팬덤 댓글화력지원 시작 드루킹 스캔들 보도 정치권 댓글 화력지원 시작 인사청문회 세월호 1 주기

37. BK & Kangmin w/ Chulalongkorn: Host-Efficient Networking Stack NICs are being equipped with more and more resources Networking stack can utilize plentiful memory of NICs 40 (best paper at APNet ‘23)

38. Traits We Look for in New Students HOT H onest O utspoken T eam player 이걸요 ? 제가요 ? 왜요 ? 41