Presentation is loading. Please wait.

Presentation is loading. Please wait.

Modelli di differenziazione delle prestazioni per il supporto del traffico LHC1 Modelli di differenziazione delle prestazioni per il supporto del traffico.

Similar presentations


Presentation on theme: "Modelli di differenziazione delle prestazioni per il supporto del traffico LHC1 Modelli di differenziazione delle prestazioni per il supporto del traffico."— Presentation transcript:

1 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC1 Modelli di differenziazione delle prestazioni per il supporto del traffico di rete LHC Tiziana.Ferrari@cnaf.infn.it On behalf of: S.Arezzini, M.Bencivenni, T.Ferrari, E.Mazzoni Workshop sul Calcolo e Reti INFN – Verso la Sfida di LHC Otranto, Jun 7 2006

2 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC2 Problem statement T1- associated T2 sites: shared network infrastructure –Optimal network bandwidth allocation (guaranteed – BGA – and max – BEA) for INFN T2 sites, considering the bursty nature of traffic produced by analysis? How to protect legacy traffic without relying on excessing overprovisioning? Solution: –Usage of IP traffic performance differentiation –Configuration of queues dedicated to specific traffic classes at potential network bottlenecks –Flow aggregation into classes of service via the Differentiated Services Code Point (6 bit, IP header)

3 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC3 What QoS model? GEANT2 and GARR: –Support of Differentiated Services technologies! –IP Premium (low delay, packet loss probability minimized)  bandwidth allocated is a “small” percentage of the overall network interface bandwidth –Less Than Best Effort (for applications which can tolerate high istantaneous packet loss)  not for TCP-based bulk data transfer  Assurate Rate service: guaranteed minimum average bandwidth to n different classes, with spare bandwidth can be re-allocated to busy queues

4 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC4 Assured Rate: an example Assumption: bandwidth of link experimenting congestion: 1 Gb/s 1.User (legacy) traffic  flows classified and assigned to dedicated queue: Guaranteed Bandwidth in range: [300, 1000] Mb/s Minimum 30% of link capacity guaranteed in case of congestion Codepoint: 000 (best-effort) 2.LHC traffic  flows classified and assigned to dedicated queue: Guaranteed Bandwidth in range: [700, 1000] Mb/s Minimum 70% of link capacity guaranteed in case of congestion Codepoint: 001 (assured-rate) 3.... And more traffic classes can be added (total max bandwidth is 100% of the link capacity on a given interface) Objectives: –Allocation of minimum guaranteed bandwidth to input/output legacy and LHC traffic classes in case of congestion –Fair distribution of link capacity in case of congestion –Possibility to get more bandwidth than the minimum guaranteed in case of spare link capacity

5 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC5 Testbed configuration (Dic 2005) CNAF: –Juniper M10 (dedicated to testing) –GigaEthernet switch Extreme Summit 400 –Two end-nodes (64 bit PCI slot network interface, 1 GEthernet), connected to the Service Challenge GigaEthernet switch –Capacity to/from GARR: 2 Gb/s (boundling of two GEthernet interfaces) PISA –Juniper M7 (production router) –Two end-nodes (64 bit PCI-X slot network interface, 1 GEthernet; 1 Fast-Ethernet interface ) –Capacity to/from GARR: 1 Gb/s

6 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC6 Network Layout GARR GARR CNAF INFN Pisa 70% 30% Juniper M10Juniper M7 1 Gb/s 2.0 Gb/s Service Challenge Network bottlenecks Users

7 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC7 Scenario 1: differentiation of incoming traffic T1  T2 GARR GARR CNAF INFN Pisa 70% 30% Juniper M10Juniper M7 1 Gb/s 2.0 Gb/s Service Challenge Users Network bottleneck  Classification and Queuing here GARR rehalm LHC subnet

8 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC8 Scenario 2: differentiation of outgoing traffic T2  T1 GARR GARR CNAF INFN Pisa 70% 30% Juniper M10 Juniper M7 1 Gb/s 2.0 Gb/s Service Challenge LHC subnet Users Network bottleneck  Classification and Queuing here INFN T2 rehalm

9 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC9 Scenario 3: differentiation of outgoing traffic T1  T2 Example GARR CNAF Pisa Juniper M10 Torino Legnaro Milano Bari 2.0 Gb/s 20% Network bottleneck  Classification and Queuing here INFN T1 rehalm LHC subnet Other Production traffic

10 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC10 Scenario 4: differentiation of incoming traffic T2  T1 GARR GARR CNAF Pisa Juniper M10 Torino Legnaro Milano Bari 2.0 Gb/s 1Gb/s Network bottleneck  Classification and Queuing here GARR rehalm

11 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC11 Differentiation of outgoing traffic (from Pisa) GARR GARR CNAF INFN Pisa SC 1 70% Juniper M10Juniper M7 1 Gb/s 2.0 Gb/s Service Challenge bottleneck Users 30% SC 2 BE1 BE2 Best effort stream 1 AR Best effort stream 2

12 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC12 Test on differentiation of outgoing traffic (1/2) Output Rate at sender [Mb/s] Input Rate At receiver [Mb/s] Transpo rt Protocol Bandwidth Distribution BE1 = 90 BE2 = 250 AR = 750 BE1 = 72 BE2 = 198 AR = 670 UDP BE1 = 8 BE2 = 21 AR = 71 BE1 = 90 BE2 = 280 AR = 700 BE1 = 67 BE2 = 208 AR = 663 UDP BE1 = 7 BE2 = 22 AR = 71 29% -Usage of Weighted Round Robin scheduling for per-class differentiation -Classification of outgoing traffic (via IP source/destination addresses) and Packet marking  transparent transport needed (no code point re-writing In transit nodes)

13 Modelli di differenziazione delle prestazioni per il supporto del traffico LHC13 Results and Conclusion Scheduling and classification/marking: good functionality, stability and performance, easy coexistence in a production router Testing with TCP streams more difficult (as expected) Scenario 1 and 2 are the most interesting... But scenario 3 and 4 getting more relevant as T1 will need to handle transfers to/from other T1s and non-associated T2s in addition to traffic from CERN (currently the LHC OPN is not fully meshed) Easy configuration on Juniper routers... But heterogeneous network layouts require extensive testing on a number of different router platforms Support of Assured Rate needed in Provider Edge routers at GARR for effective protection of incoming traffic at T1 and T2 10 GigaEthernet at T2?? –Infrastructure cost vs QoS configuration/management overhead


Download ppt "Modelli di differenziazione delle prestazioni per il supporto del traffico LHC1 Modelli di differenziazione delle prestazioni per il supporto del traffico."

Similar presentations


Ads by Google