Presentation is loading. Please wait.

Presentation is loading. Please wait.

Network speed tests CERN 14-dec-2010.

Similar presentations


Presentation on theme: "Network speed tests CERN 14-dec-2010."— Presentation transcript:

1 Network speed tests CERN 14-dec-2010

2 Hardware used 2 identical servers:
Dell PowerEdge T610 with 2x Intel xeon X5650, 24GB ram 2x PCI-Ex x8, 2x PCI-Ex x4 2x10Gb on copper cards, Broadcom chip (NetXtreme II ), PCI-Ex x8 1 switch Dell PowerConnect 6224: 24 Gb ports plus 2x10Gbase-T ports module

3 IPerf test tcp benchmark, bi-directional
do not really expect to use tcp, but still a good hw test very similar results using one-way or bi-directional transfers identical results with or without passing through the switch

4 IPerf results ~9.2 - 9.3 Gb/s with switch ~9.3 - 9.4 Gb/s without
[5] sec 1.09 GB 9.37 Gbits/sec [5] sec 10.8 GB 9.31 Gbits/sec (one direction) [4] sec 1.09 GB Gbits/sec [4] sec 10.7 GB 9.21 Gbits/sec (other direction) ~ Gb/s with switch ~ Gb/s without

5 UDP results Essentially same (disappointing) performance with
HPCBENCH tool # UDP communication test -- Mon Dec 13 23:14: # Fixed packet size unidirectional stream test # Hosts: gtktest1 (client) <--> (server) # Client UDP socket buffer size (Bytes) -- SNDBUF: RCVBUF: 65508 # Server UDP socket buffer size (Bytes) -- SNDBUF: RCVBUF: 65508 # Client IP TOS type: Default # Server IP TOS type: Default # UDP datagram (packet) size (Bytes) -- Client: 1460 Server: 1460 # Data size of each read/write (Bytes) -- Client: 1460 Server: 1460 # Message size (Bytes): 16384 # Test time (Second): # Test repeat: 1 # Network(Mbps) Local(Mbps) SentPkg(C) SentByte(C) RecvPkg(S) RecvByte(S) LostPkg LossRate Essentially same (disappointing) performance with HPCBENCH tool custom (very basic) client-server python application

6 UDP vs. TCP on PCI-Ex 4x slot
Due to wrong mounting (by Dell…), tests using the 10 GB cards in the PCI-Ex 4x slots (in place of 8x) Degraded performance due to reduced PCI-Ex bandwidth TCP UDP Mbit/s Bandwidth vs. Message size Bandwidth (average/min./max.) vs. Message size Bytes

7 More investigations needed!
UDP strange results UDP performance is the same, Gbits/sec (slightly depending on packet size), in both configurations (“wrong”: cards in the PCI-Ex 4x slots, and “correct”: cards in the PCI-Ex 8x slots) We’re hitting some kind of wall: Collisions? Lost packets? Optimization of frame-buffer? More investigations needed!

8 UDT (UDX) Just started investigating !!
UDP based application level data transfer protocol adds own reliability control and congestion control highly configurable, still looking at the possibilities not claiming we will use this protocol, but if we want reliability control udp is a candidate

9 Still to do Lots of work; just to mention:
understand if we need reliability control test behaviour with multiple 1Gb senders/one 10Gb receiver avoid memory-to-memory copy etc etc


Download ppt "Network speed tests CERN 14-dec-2010."

Similar presentations


Ads by Google