Congestion Management for Ethernet-based Lossless DataCenter Networks

Transcription

1 Congestion Management for Ethernet-based Lossless DataCenter Networks Pedro Javier Garcia 1, Jesus Escudero-Sahuquillo 1, Francisco J. Quiles 1 and Jose Duato 2 1: University of Castilla-La Mancha (UCLM) 2: Technical University València (UPV) NENDICA DCN: ICne

2 Abstract This paper describes congestion phenomena in lossless data center networks and its nega- tive consequences. It explores proposed solutions, analyzing their pros and cons to determine which are suited to the requirements of modern data centers. Conclusions identify important issues that should be addressed in the future.

3 Agenda Introduction Congestion Dynamics in DCNs Reducing In-Network and Incast Congestion Combining Congestion Management Mechanisms Conclusions

4 Agenda Introduction Congestion Dynamics in DCNs Reducing In-Network and Incast Congestion Combining Congestion Management Mechanisms Conclusions

5 Introduction On-Line Data Intensive (OLDI) Services [Congdon18] Require immediate answers to requests that are coming in at a high rate. End-user experience is highly dependent upon the system responsiveness. The network becomes a significant component of overall DC latency when congestion occurs in the network. Deadline = 250 ms Request Aggregator Deadline = 50 ms Aggregator Aggregator... Aggregator Deadline = 10 ms Worker Worker... Worker Worker Worker... Worker

6 Introduction Data-Center Networks (DCNs) Todays DCNs require a flexible fabric for carrying in a convergent way traffic from different types of applications, storage of control. Latency is a concern: Fabric design for DCNs must minimize or eliminate packet loss, provide high throughput and maintain low latency. These goals are crucial for applications of OLDI, Deep Learning, NVMe over Fabrics and the Cloudified Central Offices. However, congestion threatens these applications.

7 Introduction Why congestion isolation is needed? HoL-blocking dramatically degrades the network performance (e.g. PFC has not enough granularity and there is no congested flow identification) [Garcia05]. Classical e2e congestion control for lossless networks is difficult to tune, reacts slowly, and may introduce oscillations and instability [Escudero11]. Network Throughput (normalized) HS starts HS = traffic injected to Hot Spot destination HS ends 1Q ITh VOQnet 0 1e+06 2e+06 3e+06 4e+06 5e+06 Time (nanoseconds) 64-node CLOS network, 4 hot-spots

8 Introduction Why congestion isolation is needed? Src. A 33% Sw. 1 33% Sw. 5 Congested flows (Dst. X) Non-congested flows (Dst. Y) 33% Non-congested flows (Dst. Z) Src. B 33% Sw. 2 66% Sw. 6 33% Sw % Dst. X 33% 33% Src. C Src. D 33% Sw. 3 33% Sw. 7 33% 66% Sw. 9 Dst. Y 33% Dst. Z Src. E 33% Sw. 4 33% High-Order HoL-blocking Low-Order HoL-blocking 33 % Sending 33 % Stopped 33 % Sending 33 % Stopped 33 % Sending 33 % Sending

9 Introduction Why congestion isolation is needed? We need a congestion isolation (CI) mechanism that reacts quickly when transient congestion situations appear, preventing network performance degradation caused by the HoL blocking. We want a CI mechanism that complements other technologies available in the DCNs, so that CI improves their performance, while the others reduce the CI complexity.

10 Agenda Introduction Congestion Dynamics in DCNs Reducing In-Network and Incast Congestion Combining Congestion Management Mechanisms Conclusions

11 Congestion Dynamics in DCNs Appearance of Congestion Congestion Congestion Injection rate at 100% of the link bandwidth (full rate) Injection rate at 100% of the link bandwidth (full rate) Speedup = 1 Speedup = 2 Congestion (t0+t) Congestion (t0) Congestion (t0) Congestion (t0+t) Injection rate at 100% of the link bandwidth (full rate) Injection rate at 100% of the link bandwidth (full rate) Speedup = 2 Speedup = 1.5

12 Congestion Dynamics in DCNs Growth of Congestion Trees (from root to leaves) Switch 1 Switch 3 Switch speedup = 1.5 Packet flows Congestion point Switch 5 Switch 2 Switch 4

13 Congestion Dynamics in DCNs Growth of Congestion Trees (from leaves to root) Switch speedup = 1.5 Packet flows Congestion point Switch 1 Switch 5 Switch 2 Switch 7 Switch 3 Switch 6 Switch 4

14 Congestion Dynamics in DCNs Growth of Congestion Trees (Roots movement) Switch speedup = 1.5 Packet flows (start) Packet flows (after) Congestion point Switch 1 Switch 1 Switch 3 Switch 3 Switch 2 Switch 2

15 Congestion Dynamics in DCNs Growth of Congestion Trees (in-network roots) Switch 1 Switch 5 Switch 2 Switch 7 Switch 8 X Y Switch 3 Switch 6 Switch speedup = 1.5 Packet flows addressed to X Packet flows addressed to Y Congestion point Switch 4

16 Congestion Dynamics in DCNs Growth of Congestion Trees (Overlapping) X Switch 1 Switch 4 Switch 8 Switch 2 Switch 5 Switch 7 Y Switch 3 Switch 6 Switch speedup = 1.5 Packet flows addressed to X Packet flows addressed to Y Congestion point Switch 9

17 Congestion Dynamics in DCNs Growth of Congestion Trees (Vanishing) Switch speedup = 1.5 Permanent packet flows Packet flows disappearing first Congestion point first appeared in the switch Switch 1 Switch 1 Switch 3 Switch 3 Switch 2 Switch 2

18 Agenda Introduction Congestion Dynamics in DCNs Reducing In-Network and Incast Congestion Combining Congestion Management Mechanisms Conclusions

19 Reducing Congestion Incast congestion reduction - ECMP

20 Reducing Congestion In-network congestion reduction - ECN X Switch 1 Switch 4 Switch 8 Switch 2 Switch 5 Switch 7 Y Switch 3 Switch 6 Switch speedup = 1.5 Packet flows addressed to X Packet flows addressed to Y Victim flow Congestion point Switch 9

21 Reducing Congestion Limitations of current technologies [Escudero19] These technologies may work together to eliminate loss in the cloud data center network. Load-balancing and destination scheduling are end-toend solutions incurring in the RTT delays when congestion appear. However, there is no time for loss in the network due to congestion and congestion trees grow very quickly. Transient congestion may still produce HoL blocking that leads to increase latency, lower throughput and buffers overflow, significantly degrading performance. Even using these mechanisms, we still need something to deal with HOL Blocking locally and fast.

22 Agenda Introduction Congestion Dynamics in DCNs Reducing In-Network and Incast Congestion Combining Congestion Management Mechanisms Conclusions

23 Combining Congestion Management Mechanisms CI is needed to react locally and very fast to immediately eliminate HoL blocking. Previous technologies reduce the use of PFC and ECN, but their closed- and open-loop approach cause delays still happening. Congestion trees appear suddenly, are difficult to predict (even worse when load balancing is applied) and grow quickly. New techniques can be applied in combination to the previous technologies, improving their behavior.

24 Combining Congestion Management Mechanisms Dynamic Virtual Lanes (DVL) Switch A Switch B P1 CFQ ncfq CFQ ncfq P3 CIP P1 CFQ ncfq CFQ ncfq P3 Congestion Root P2 CFQ P4 P2 CFQ P4 ncfq Legend Output port requested by the packet on top. Congestion root. Congestion Isolation Packets (CIP). Packets from congested flows. Packets from non-congested flows. ncfq

25 Agenda Introduction Congestion Dynamics in DCNs Reducing In-Network and Incast Congestion Combining Congestion Management Mechanisms Conclusions

26 References [Duato03] J. Duato, S. Yalamanchili, and L. M. Ni, Interconnection Networks: An Engineering Approach. San Francisco, CA, USA: Morgan Kaufmann Publishers, [Garcia05] P. J. Garcia, J. Flich, J. Duato, I. Johnson, F. J. Quiles, and F. Naven, Dynamic Evolution of Congestion Trees: Analysis and Impact on Switch Architecture, in High Performance Embedded Architectures and Compilers, ser. Lecture Notes in Computer Science. Springer, Berlin, Heidelberg, Nov. 2005, pp [Congdon18] Paul Congdon, IEEE 802 Nendica Report: The Lossless Network for Data Centers, IEEE-SA Industry Connections White Paper, August [Leiserson85] C. E. Leiserson, Fat-trees: Universal networks for hardware-efficient supercomputing, IEEE Transactions on Computers, vol. C-34, pp , Oct [Escudero11] Jesús Escudero-Sahuquillo, Ernst Gunnar Gran, Pedro Javier García, Jose Flich, Tor Skeie, Olav Lysne, Francisco J. Quiles, José Duato: Combining Congested-Flow Isolation and Injection Throttling in HPC Interconnection Networks. ICPP 2011: [Escudero19] Jesús Escudero-Sahuquillo, Pedro Javier García, Francisco J. Quiles, José Duato: P802.1Qcz interworking with otherdata center technologies. IEEE Plenary Meeting, San Diego, CA, USA July 8, 2018 (cz-escudero-sahuquillo-ci-internetworking-0718-v1.pdf)