fla lashnet t fla lash netw twork ork sta tack c k co de
play

Fla lashNet: t: Fla lash/Netw twork ork Sta tack C k Co-De - PowerPoint PPT Presentation

May 25, 2023 •124 likes •605 views

Fla lashNet: t: Fla lash/Netw twork ork Sta tack C k Co-De Design ign Animesh Trivedi, Nikolas Ioannou, Bernard Metzler, Patrick Stuedi, Jonas Pfefferle, Ioannis Koltsidas, Kornilios Kourtis, and Thomas R. Gross IBM Research and ETH

  1. Fla lashNet: t: Fla lash/Netw twork ork Sta tack C k Co-De Design ign Animesh Trivedi, Nikolas Ioannou, Bernard Metzler, Patrick Stuedi, Jonas Pfefferle, Ioannis Koltsidas, Kornilios Kourtis, and Thomas R. Gross IBM Research and ETH Zurich, Switzerland

  2. Modern Distributed Systems  data intensive  run on 100-1000s of servers  performance depends upon both network and storage SYSTOR 2017, Haifa 2

  3. Modern Distributed Systems  performance depends upon both network and storage SYSTOR 2017, Haifa 3

  4. Modern Distributed Systems ● StackMap: Low-Latency Networking with the OS Stack and Dedicated NICs, USENIX'16 ● Network Stack Specialization for Performance, SIGCOMM'14 ● mTCP: A Highly Scalable User-level TCP Stack for Multicore Systems, NSDI'14 ● MegaPipe: A New Programming Interface for Scalable Network I/O, OSDI'12  performance depends upon ● … both network and storage SYSTOR 2017, Haifa 4

  5. Modern Distributed Systems ● StackMap: Low-Latency Networking with the OS Stack and Dedicated NICs, USENIX'16 ● Network Stack Specialization for Performance, SIGCOMM'14 ● mTCP: A Highly Scalable User-level TCP Stack for Multicore Systems, NSDI'14 ● MegaPipe: A New Programming Interface for Scalable Network I/O, OSDI'12  performance depends upon ● … both network and storage ● NVMeDirect: A User-space I/O Framework for Application-specific Optimization on NVMe SSDs, HotStorage'16 ● OS I/O Path Optimizations for Flash Solid-state Drives, USENIX'14 ● Linux Block IO: Introducing Multi-queue SSD Access on Multi-core Systems, SYSTOR'13 ● When Poll is Better Than Interrupt, FAST'12 ● … SYSTOR 2017, Haifa 5

  6. Modern Distributed Systems ● StackMap: Low-Latency Networking with the OS Stack and Dedicated NICs, USENIX'16 ● Network Stack Specialization for Performance, SIGCOMM'14 ● mTCP: A Highly Scalable User-level TCP Stack for Multicore Systems, NSDI'14 ● MegaPipe: A New Programming Interface for Scalable Network I/O, OSDI'12  performance depends upon ● … both network and storage ● NVMeDirect: A User-space I/O Framework for Application-specific Optimization on NVMe SSDs, HotStorage'16 ● OS I/O Path Optimizations for Flash Solid-state Drives, USENIX'14 ● Linux Block IO: Introducing Multi-queue SSD Access on Multi-core Systems, SYSTOR'13 ● When Poll is Better Than Interrupt, FAST'12 ● … SYSTOR 2017, Haifa 6

  7. Modern Distributed Systems ● StackMap: Low-Latency Networking with the OS Stack and Dedicated NICs, USENIX'16 ● Network Stack Specialization for Performance, SIGCOMM'14 ● mTCP: A Highly Scalable User-level TCP Stack for Multicore Systems, NSDI'14 ● MegaPipe: A New Programming Interface for Scalable Network I/O, OSDI'12  performance depends upon ● … both network and storage ● NVMeDirect: A User-space I/O Framework for Application-specific Optimization on NVMe SSDs, HotStorage'16 ● OS I/O Path Optimizations for Flash Solid-state Drives, USENIX'14 ● Linux Block IO: Introducing Multi-queue SSD Access on Multi-core Systems, SYSTOR'13 ● When Poll is Better Than Interrupt, FAST'12 ● … SYSTOR 2017, Haifa 7

  8. The Cost of the Gap 1400 1200 1000 800 K IOPS 600 400 200 0 spec. block IO netperf iSCSI KV NFS HDFS SYSTOR 2017, Haifa 8

  9. The Cost of the Gap 1400 1200 1000 800 K IOPS 600 400 200 0 spec. block IO netperf iSCSI KV NFS HDFS SYSTOR 2017, Haifa 9

  10. The Cost of the Gap 1400 1200 1000 800 K IOPS 600 400 200 0 spec. block IO netperf iSCSI KV NFS HDFS SYSTOR 2017, Haifa 10

  11. The Reason for the Gap client flash storage SYSTOR 2017, Haifa 11

  12. The Reason for the Gap 1 request 2 request processing client flash 3 response storage SYSTOR 2017, Haifa 12

  13. The Reason for the Gap 1 request 2 request processing client flash 3 response storage performance = network IO + server time + storage IO SYSTOR 2017, Haifa 13

  14. The Reason for the Gap 1 request 2 request processing client flash 3 response storage performance = network IO + server time + storage IO application involvement scheduling fs lookups and overheads ... SYSTOR 2017, Haifa 14

  15. A Detailed Look: send usersp userspace kern ernel 1. TCP/IP processing SYSTOR 2017, Haifa 15

  16. A Detailed Look: send 3. receive request usersp userspace kern ernel 2. receive processing 1. TCP/IP processing SYSTOR 2017, Haifa 16

  17. A Detailed Look: send 3. receive request usersp userspace kern ernel 4. fs translation 2. receive processing 5. block I/O 1. TCP/IP processing SYSTOR 2017, Haifa 17

  18. A Detailed Look: send 3. receive 7. response request transmission userspace usersp kern ernel 4. fs translation 2. receive processing 5. block I/O 1. TCP/IP 6. flash I/O processing completion SYSTOR 2017, Haifa 18

  19. A Detailed Look: send 3. receive 7. response request transmission userspace usersp kern ernel 4. fs translation 8. send 2. receive processing 5. block I/O 1. TCP/IP 6. flash I/O 9.TX done processing completion SYSTOR 2017, Haifa 19

  20. A Detailed Look: sendfjle 3. receive 7. response request transmission userspace usersp kern ernel 4. fs translation 8. send 2. receive processing 5. block I/O 1. TCP/IP 6. flash I/O 9.TX done processing completion SYSTOR 2017, Haifa 20

  21. A Detailed Look: sendfjle 3. receive 7. response request transmission userspace usersp kern ernel 4. fs translation 8. send 2. receive processing 5. block I/O 1. TCP/IP 6. flash I/O 9.TX done processing completion SYSTOR 2017, Haifa 21

  22. The FlashNet Approach 3. receive request usersp userspace kern ernel 4. fs translation 2. receive processing 5. block I/O 1. TCP/IP processing SYSTOR 2017, Haifa 22

  23. The FlashNet Approach eliminate application involvement 3. receive request userspace usersp kern ernel 4. fs translation 2. receive processing 5. block I/O 1. TCP/IP processing SYSTOR 2017, Haifa 23

  24. The FlashNet Approach eliminate application involvement 3. receive request usersp userspace reduce file kern ernel 4. fs translation system overheads 2. receive processing 5. block I/O 1. TCP/IP processing SYSTOR 2017, Haifa 24

  25. The FlashNet Approach eliminate application involvement 3. receive request usersp userspace reduce file kern ernel 4. fs translation system overheads 2. receive processing 5. block I/O enable direct network and storage interaction 1. TCP/IP processing SYSTOR 2017, Haifa 25

  26. The FlashNet Approach eliminate application involvement RDMA usersp userspace reduce file kern ernel 4. fs translation system overheads 2. RDMA processing 5. block I/O enable direct network and storage interaction 1. TCP/IP processing SYSTOR 2017, Haifa 26

  27. The FlashNet Approach eliminate application involvement RDMA usersp userspace reduce file kern ernel simple fs layout system overheads 2. RDMA processing 5. block I/O enable direct network and storage interaction 1. TCP/IP processing SYSTOR 2017, Haifa 27

  28. The FlashNet Approach eliminate application involvement RDMA usersp userspace reduce file kern ernel simple fs layout system overheads 2. RDMA processing use VM 5. block I/O enable direct network and storage interaction 1. TCP/IP processing SYSTOR 2017, Haifa 28

  29. The FlashNet Approach usersp userspace 2. RDMA v to L 6. response kern ernel a B A transmission 2. RDMA processing 3. block I/O 1. TCP/IP 6. flash I/O 7.TX done processing completion SYSTOR 2017, Haifa 29

  30. FlashNet: A Co-Designed Network and Storage Stack 64-bit LBA space flash controller ● flash virtualization ● I/O management ● ... SYSTOR 2017, Haifa 30

  31. FlashNet: A Co-Designed Network and Storage Stack ● contigious file allocation ● supporting m & local file I/O m a p ● … ContigFS flash controller SYSTOR 2017, Haifa 31

  32. FlashNet: A Co-Designed Network and Storage Stack ● lazy RDMA pinning ● resolving flash & file addresses ● … ContigFS RDMA controller flash controller SYSTOR 2017, Haifa 32

  33. FlashNet: A Co-Designed Network and Storage Stack FlashNet I/O stack ContigFS RDMA controller flash controller SYSTOR 2017, Haifa 33

  34. FlashNet: A Co-Designed Network and Storage Stack file server application virtual address LBA ContigFS RDMA STag PBA controller flash controller network control setup expanding to storage SYSTOR 2017, Haifa 34

  35. FlashNet: A Co-Designed Network and Storage Stack file server application virtual address LBA ContigFS RDMA STag PBA controller flash controller network control setup expanding to storage data path from a flash device to a client buffer SYSTOR 2017, Haifa 35

  36. FlashNet: A Co-Designed Network and Storage Stack client server application nfs ContigFS RDMA iSCSI controller flash RNIC controller SKB network control setup expanding to storage data path from a flash device to a client buffer SYSTOR 2017, Haifa 36

  37. Performance Evaluation How efficient is FlashNet's IO path? Does it help with applications? ...more in the paper 9-machine cluster testbed CPU : dual socket E5-2690, 2.9 GHz, 16 cores DRAM : 256 GB, DDR3 1600 MHz NIC : 40Gbit/s Ethernet 3xNVMe Flash : 6.6 GB/sec (read), 2.7 GB/sec (write) peak 4kB read IOPS: 1.3 M SYSTOR 2017, Haifa 37

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

NE NETW TWORK ORK Presented by: Eng. S. M. Kinoti Ag. DG -KURA TH FEBR UB ON 8 TH Meeti

NE NETW TWORK ORK Presented by: Eng. S. M. Kinoti Ag. DG -KURA TH FEBR UB ON 8 TH Meeti

1 UPGRADING RADING OF NAI AIROB ROBI I URB RBAN AN RO ROAD AD NE NETW TWORK ORK Presented by: Eng. S. M. Kinoti Ag. DG -KURA TH FEBR UB ON 8 TH Meeti ting g at NAIR IROB OBI I CLUB BRUAR ARY, 2017 17 OUTL TLINE INE 2

510 views • 30 slides
cryptography in ITU-T and ISO/IEC Hao Hao Qi Qin* CAS Q S Quant ntum Netw twork rk Co., .,

cryptography in ITU-T and ISO/IEC Hao Hao Qi Qin* CAS Q S Quant ntum Netw twork rk Co., .,

QCrypt 2020 Industry session August 12, 2020 E-meeting Standardization of quantum cryptography in ITU-T and ISO/IEC Hao Hao Qi Qin* CAS Q S Quant ntum Netw twork rk Co., ., Ltd td. *qinhao@casquantumnet.com Quantum key distribution:

161 views • 13 slides
Ultra L Low Power er Infer eren ence a e at the v very e edge o of the n netw twork

Ultra L Low Power er Infer eren ence a e at the v very e edge o of the n netw twork

Ultra L Low Power er Infer eren ence a e at the v very e edge o of the n netw twork Tiny ML Summit March 20-21 2019 Sunnyvale Eric Flamand, CTO & CoFounder of Greenwaves Technologies Wh Who a o are w we? French based startup

644 views • 25 slides
Trackin acking g th the e Rus ussi sian n Bus usin ines ess s Netw twork rk (RBN)

Trackin acking g th the e Rus ussi sian n Bus usin ines ess s Netw twork rk (RBN)

Trackin acking g th the e Rus ussi sian n Bus usin ines ess s Netw twork rk (RBN) Knowing and understanding the RBN is a useful objective, but surely the main goal is to stop them? RBNexploit.blogspot.com Cambridge University

554 views • 34 slides
How to Displa lay Multiple Ad Netwo works? What is an Ad Netw twork? Popular Ad network

How to Displa lay Multiple Ad Netwo works? What is an Ad Netw twork? Popular Ad network

How to Displa lay Multiple Ad Netwo works? What is an Ad Netw twork? Popular Ad network rks How to integrate Ad Ad Networks Revenu nue Models There are three differ ferent ways app developers make re revenue from their apps

1.04k views • 40 slides
Fi Financing nancing the gas gas tr trans ansmis mission sion netw twork ork Analyzing

Fi Financing nancing the gas gas tr trans ansmis mission sion netw twork ork Analyzing

Fi Financing nancing the gas gas tr trans ansmis mission sion netw twork ork Analyzing the effects of duration dependent multipliers Eren am | Dominic minic Lencz ncz | Simon Schulte | ENERDAY - Dresden | 12.04.2019 Agenda da

208 views • 16 slides
New York Ne ork & Ne New Jersey UpdateJu July y 2018 Ne Netw twork rk Membe mber

New York Ne ork & Ne New Jersey UpdateJu July y 2018 Ne Netw twork rk Membe mber

New York Ne ork & Ne New Jersey UpdateJu July y 2018 Ne Netw twork rk Membe mber r Webi bina nar Hosted by: William OHearn, Communications/Outreach Manager, NJ and NY Elizabeth Barminski, Working Group Coordinator

703 views • 21 slides
Maxi ximis isin ing customer valu lue fr from th the netw twork in in a hig igh-DER fu

Maxi ximis isin ing customer valu lue fr from th the netw twork in in a hig igh-DER fu

1 SA Power Networks Maxi ximis isin ing customer valu lue fr from th the netw twork in in a hig igh-DER fu futu ture AEMC / ARENA Regulatory DEIP dive, 6 th June 2019 The fu futu ture of f th the dis istribution network

283 views • 16 slides
Artificial Artificial neur neural al ne netw twork ork alg algorithm rithms fo for Heal

Artificial Artificial neur neural al ne netw twork ork alg algorithm rithms fo for Heal

Artificial Artificial neur neural al ne netw twork ork alg algorithm rithms fo for Heal Healthcar care Decision Decisi on Syste Systems and and Di Diagnos agnostic ic Im Image Processing ocessing T A H M I N A Z E B I N R E S E A R C H F E

649 views • 27 slides
Building the gigabit society Eelco Blok KPN at a glance Mobile 4G netw twork rk with 99%

Building the gigabit society Eelco Blok KPN at a glance Mobile 4G netw twork rk with 99%

BoR (16) 199 Building the gigabit society Eelco Blok KPN at a glance Mobile 4G netw twork rk with 99% 99% 99% geographi hic and popula lation ion coverage World ld first st natio ionwid ide LoRa network ork for 99% 99%

325 views • 8 slides
Multi lti-le level l Attentio ion Netw twork rk usin sing Text xt, Audio io and Vide

Multi lti-le level l Attentio ion Netw twork rk usin sing Text xt, Audio io and Vide

Multi lti-le level l Attentio ion Netw twork rk usin sing Text xt, Audio io and Vide ideo for r Depressio ion Predic ictio ion Anupama Ray Siddharth Kumar, Rutvik Reddy, Prerana Mukherjee Ritu Garg IBM Research, India

634 views • 19 slides
Approach to Coding Reform A Smart Gr A Smart Growth th Ne Netw twork W ork Web ebin inar

Approach to Coding Reform A Smart Gr A Smart Growth th Ne Netw twork W ork Web ebin inar

Enabling Better Places: An Incremental Approach to Coding Reform A Smart Gr A Smart Growth th Ne Netw twork W ork Web ebin inar ar July July 17, 17, 2019 2019 Lyn ynn Rich Richar ards: Con s: Congr gress f ess for th or the Ne

878 views • 32 slides
The he Capac apacity ity of of th the Ag e Aging ng Service vices s Netw twork: ork:

The he Capac apacity ity of of th the Ag e Aging ng Service vices s Netw twork: ork:

The he Capac apacity ity of of th the Ag e Aging ng Service vices s Netw twork: ork: Top op 10 Thi hings ngs to to Kno now w fo for Reautho uthorization rization 1 1965 1965 1972 1972 1973 1973 1975 1975 OAA,

771 views • 40 slides
Con Convol oluti tion onal Neural Netw twork orks Presented by Tristan Maidment Adapted

Con Convol oluti tion onal Neural Netw twork orks Presented by Tristan Maidment Adapted

Con Convol oluti tion onal Neural Netw twork orks Presented by Tristan Maidment Adapted from Ke Yus Slides Ou Outlin line Neural Network recap Building blocks of CNNs Architecture of CNNs Visualizing and understanding

808 views • 70 slides
Lectu ture 6 6 reca recap Prof. Leal-Taix and Prof. Niessner 1 Ne Neural Ne Netw twork

Lectu ture 6 6 reca recap Prof. Leal-Taix and Prof. Niessner 1 Ne Neural Ne Netw twork

Lectu ture 6 6 reca recap Prof. Leal-Taix and Prof. Niessner 1 Ne Neural Ne Netw twork Width Depth Prof. Leal-Taix and Prof. Niessner 2 Gr Gradi dient De Descent fo for Neural Netwo works ks " =: ! " ' " (

1.42k views • 79 slides
in Regards to Tack. David Branscome Jr. Operations Branscome Paving Company Effects on Tack

in Regards to Tack. David Branscome Jr. Operations Branscome Paving Company Effects on Tack

Problems Contractors Face in Regards to Tack. David Branscome Jr. Operations Branscome Paving Company Effects on Tack Due to Short Time Frame Reduced hours of actual paving lay down time Keeping quality while trying to get the

457 views • 11 slides
Sequencing, and IO We said functional programming is about calculating expressions Bjrn Lisper

Sequencing, and IO We said functional programming is about calculating expressions Bjrn Lisper

Sequencing in F# Sequencing, and IO We said functional programming is about calculating expressions Bjrn Lisper Simple way of interacting: type an expression, obtain the calculated result School of Innovation, Design, and Engineering

340 views • 8 slides
Amdahls Law How is system performance altered when some component is changed? Example 1:

Amdahls Law How is system performance altered when some component is changed? Example 1:

Amdahls Law How is system performance altered when some component is changed? Example 1: Program execution time is made up of 75% CPU time and 25% I/O time. Which is the better enhancement: (a) Increasing the CPU speed by 50% or (b)

342 views • 12 slides
1. Introduction @ MaxChamb - chamberm.github.io There is an urgent need for a paradigm shift from

1. Introduction @ MaxChamb - chamberm.github.io There is an urgent need for a paradigm shift from

Tractometry-based Anomaly Detection for Single-subject White Matter Analysis Chamberland et al. MIDL 2020 1. Introduction @ MaxChamb - chamberm.github.io There is an urgent need for a paradigm shift from group-wise comparisons 1 (N vs M) to

92 views • 5 slides
1 2 3 4 printChar 5 printChar 6 printChar xs = [printChar a, printChar b]

1 2 3 4 printChar 5 printChar 6 printChar xs = [printChar a, printChar b]

1 2 3 4 printChar 5 printChar 6 printChar xs = [printChar a, printChar b] printChar 7 8 9 type IO = World -> World type IO a = World -> (a, World) b -> World -> (a, World) 10 IO a IO Char IO ()

858 views • 34 slides
Input/Output CS 351: Systems Programming Michael Saelee <lee@iit.edu> Computer Science

Input/Output CS 351: Systems Programming Michael Saelee <lee@iit.edu> Computer Science

Input/Output CS 351: Systems Programming Michael Saelee <lee@iit.edu> Computer Science Science CPU Memory I/O shared network disk terminal printer memory Computer Science Science shared network disk terminal

1.48k views • 109 slides
Draft RTP Profile for Speex Audio Authors: Greg Herlein, Jean-Marc Valin, Simon Morlat, Roger

Draft RTP Profile for Speex Audio Authors: Greg Herlein, Jean-Marc Valin, Simon Morlat, Roger

Draft RTP Profile for Speex Audio Authors: Greg Herlein, Jean-Marc Valin, Simon Morlat, Roger Hardiman, Phil Kerr IETF Audio/Video Transport Working Group July 14th, 2003, Vienna, Austria draft-herlein-speex-rtp-profile-01.txt About Speex

297 views • 6 slides
Higher Order Combinators for Join Patterns using STM Satnam Singh, Microsoft TRANSACT 2006

Higher Order Combinators for Join Patterns using STM Satnam Singh, Microsoft TRANSACT 2006

Higher Order Combinators for Join Patterns using STM Satnam Singh, Microsoft TRANSACT 2006 Overview Specifically: encoding an existing concurrency idiom with STM very straightforward nothing clever More generally: what kind of

249 views • 24 slides
Principles of Software Construction: I/O and reflection Josh Bloch Charlie Garrod School of

Principles of Software Construction: I/O and reflection Josh Bloch Charlie Garrod School of

Principles of Software Construction: I/O and reflection Josh Bloch Charlie Garrod School of Computer Science 15-214 1 Administrivia Homework 4c due tonight Homework 5 coming out tomorrow Midterm next Thursday in class Midterm

592 views • 28 slides

More recommend