Radon: Network QoS. Andrew Shewmaker Carlos Maltzahn Katia Obraczka Scott Brandt Ivo Jimenez. UC Santa Cruz Graduate Studies in Computer Science

Transcription

1 Radon: Network QoS Andrew Shewmaker Carlos Maltzahn Katia Obraczka Scott Brandt Ivo Jimenez UC Santa Cruz Graduate Studies in Computer Science Oct 22nd, 2013 This research was made possible by LANL, Cisco, and Google 1 / 20

2 Radon Targets Highly Contended, Parallel Bottlenecks The purpose of Radon (Resource Allocation and Dispatching On Networks) is to maximize utilization of network gateways where bandwidth is precious 1 storage networks exascale I/O cloud SLAs datacenter provisioning WAN gateways 2 M 2 / 20

3 How Much QoS Enforcement At Each Layer? host app transport protocol app transport protocol globally by a centralized traffic controller per flow by the host queueing disciplines per application by the transport layer per application by the application 1 2 M switch tx port bottleneck queue queueing discipline Global Ethernet SDN Controller / Infiniband Subnet Manager 3 / 20

4 Radon s Enforcement Choices host app transport protocol app transport protocol global route scheduling by a central controller per flow rate limiting by the host per application rate limiting and congestion avoidance by the transport layer 1 2 M switch tx port bottleneck queue queueing discipline Global Ethernet SDN Controller / Infiniband Subnet Manager 4 / 20

5 Don t Use Priorities If You Know What You Want RAD (utilization, period) - how much? how often? enable arbitrarily fine-grained QoS keep meanings consistent in a dynamic environment reason across global, flow, and application layers Priority RAD reservation A 2? (05, 1s) B 3 (033, 075s) C 1? (014, 025s) / 20

6 SDN Controller - Admission control Admit if the sum of all utilizations does not exceed M, for M routes Priority RAD reservation A 2? (05, 1s) B 3 (033, 075s) C 1? (014, 025s) / 20

7 SDN Controller - Route Management Reduction to Uniprocessor Unipath (RUN) Regnier, P, et al Multiprocessor scheduling by reduction to uniprocessor: an original optimal approach RTSS 2011 (Best Paper Award) reduce N flows on M routes schedule single path with well-known algorithms transforms back to multipath schedule 1 2 M 7 / 20

8 Existing Host Queueing qdiscs provide strict rate limiting rate-limited for any measured interval not work conserving cpu/numa mappings head-of-line blocking multiqueue scheduler avoids head-of-line blocking round-robin is fair, but not QoS-aware network classifier cgroup tags packets for specific qdiscs 8 / 20

9 Proposed Host Queueing Reduction to Uniprocessor Uniqueue (RUN) limits migrations between queues prevents head-of-line blocking short term: control groups + strict rate limiting qdiscs long term: control groups + flexible RAD-based qdisc 9 / 20

10 Transport Layer Quotes from the 2013 ACM Symposium on Cloud Computing: Whatever the major protocol will be, it will be called TCP Buffering don t do it TCP is here to stay, but it needs to be improved work conserving we need something to use up unused bandwidth knows about congestion it needs to know more needs to be aware of RAD reservations 10 / 20

11 CUBIC TCP Experiment - Setup Mininet Setup Server Client Switch 100 mbit/s, 52 ms delay netem Switch 100 mbit htb Client This setup replicates Figure 9 from the Hystart Technical report 11 / 20

12 CUBIC TCP Experiment - Congestion Window Typical cubic congestion responses cwnd, ssthresh (segments) h1 snd_cwnd h2 snd_cwnd h1 snd_ssthresh h2 snd_ssthresh Time (s) 12 / 20

13 CUBIC TCP - Bottleneck Queue & Round Trip Times Standing/filling/dropping bottleneck queue clear RTT trends 13 / 20

14 Next Generation TCP: Hypercubic To Do: use Hystart detection as general congestion signal enhance Hystart detection add cgroup and RAD reservation awareness 14 / 20

15 CUBIC TCP Experiment Hypercubic bottleneck queue < 500 and prevents packet loss 15 / 20

16 Going Forward Reduce To Unipath (RUN) Reduce To Uniqueue (RUN) fix TCP (Run! Just kidding) I m looking for partners willing to try these ideas on a large scale testbed 16 / 20

17 Thanks! more on RAD Dynamic Integrated Scheduling of Hard Real-Time, Soft Real-Time and Non-Real-Time Processes original RUN paper by Regnier, P, et al (RTSS11 Best Paper Award) Multiprocessor scheduling by reduction to uniprocessor: an original optimal approach Run, Fatboy, Run: Applying the Reduction to Uniprocessor Algorithm to Other Wide Resources tech report 17 / 20

18 Backup Slides 18 / 20

19 Hypercubic Congestion Window cwnd, ssthresh (segments) h1 snd_cwnd h2 snd_cwnd h1 snd_ssthresh h2 snd_ssthresh Time (s) 19 / 20

20 Round Trip Times Comparison CUBIC Hypercubic 20 / 20