1 / 17

Implementation of a Network Processor

Implementation of a Network Processor. G.Grosdidier , P.Matricon , K.Petrov LAL/IN2P3/CNRS & SPhN /IRFU/CEA ( PetaQCD project , 19/10/2011) Confidential. Block Design. Multicore Processors. Mem . Bus. PCI Bus. RAM. GPU. GPU. Infiniband. DRAM. Mem ory Bus. Multicore

beau-woodard
Télécharger la présentation

Implementation of a Network Processor

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Implementationof aNetwork Processor G.Grosdidier, P.Matricon, K.Petrov LAL/IN2P3/CNRS & SPhN/IRFU/CEA (PetaQCDproject, 19/10/2011) Confidential

  2. Block Design Multicore Processors Mem. Bus PCI Bus RAM GPU GPU Infiniband

  3. DRAM Memory Bus Multicore Processors PCI Bus G D D R G D D R G D D R GPU GPU GPU Network

  4. Required Multicore Processors (Intel, AMD ?) Mem. Bus 8GB/s r/w mem bus 8GB/sec both ways 6GB++/socket RAM RAM PCI Bus 6GB++/card RAM GPU Nvidia2090 ? GPU GPU 8x 2GB/sec both ways Infiniband ??

  5. Current Multi-core Processors (Intel, AMD ?) ~10GB/s r/w Mem. Bus 8GB/s r/w mem bus 8GB/sec both ways 6GB++/socket RAM RAM PCI Bus ~5 GB/sec both ways 6GB++/card GDDR GPU Nvidia2090 ? GPU GPU 8x 2GB/sec both ways Nothing 0GB/sec

  6. Current GPU-Direct • The performances have been measured in INRIA-Bordeaux • This is obviously preliminary • They got 50 μsec latency ‘one-way’ • This is by far too large for our requirements • What else ? • Lets assume we want these transfer durations to be negligible compared to main computation times, ie at least 10x faster

  7. Next step ? • What can be done to enhance current setup ? • There seems to be a link already between the GPU chip and the HCA one • Green arrow below • Why not to use a straight route thru this link towards the other GPU ? • Red path

  8. In the end • Replace the Infiniband HCA with Custom Network Processors • Better suited to our 8 NTN specific connectivity toplogy • Get rid of any network switches • Does not avoid transit through PCIe though (latencies …) • Increases the PCIe bottleneck on GPU exit Custom Network Processors

  9. Inserting a Network Processor SystemRAM SystemRAM Memory Bus Memory Bus Multicore Processors CPU CPU QPI QPI QPI QPI Chipset Chipset Chipset Chipset Infiniband Infiniband IB HCA IB HCA PCIe x16 PCIe x16 GPU GPU GPU GPU Custom Protocol Network Processor Network Processor Network Processor Network Processor 8 Nearest Topological Neighbours 8 N.T.N. 8 N.T.N. 8 N.T.N.

  10. Simplest case : 1 socket and 1 Chipset DRAM CPU QDR/ EDR/ FDR HCA/IB DMI Chipset 8 N.T.N. (optical links) PCIe Controller NP1 FPGA 16x 8X 8X NP2 FPGA GPU GDDR

  11. Data routes DRAM CPU QDR/ EDR/ FDR HCA/IB DMI Chipset 8 N.T.N. (optical links) PCIe Controller NP1 FPGA 16x 8X 8X NP2 FPGA GPU GDDR

  12. 3 cards DRAM CPU QDR/ EDR/ FDR HCA/IB DMI Chipset 8 N.T.N. (optical links) PCIe Controller NP1 FPGA 16x 8X 8X NP2 FPGA GPU GDDR

  13. Sandy Bridge (expected)

  14. Sandy Bridge setup, 2 sockets

  15. Sandy Bridge case(?) DRAM Infiniband QDR/EDR/FDR CPU (multicore socket) DMI HCA/IB Chipset 4X 8 N.T.N. (optical links) PCIe Controller NP1 FPGA 16x 8X 8X NP2 FPGA GPU GDDR

  16. Infiniband QDR/EDR/FDR NP1 8 N.T.N. (optical links) HCA/IB DRAM CPU (multicore socket) 4X DMI Chipset UC1 NP2 PCIe Controller Network Processor NP0 (FPGA) 8 N.T.N. (optical links) 8 N.T.N. (optical links) 16x 8X 8X 8x 8x 8x GPU GDDR 8x 8x 8x UC2 UC0 NP3 8 N.T.N. (optical links) UC3

  17. 8X 8X 8X 8X 8x 8x

More Related