Upload
arnold-simmons
View
216
Download
0
Embed Size (px)
Citation preview
Department of Particle Physics & Astrophysics
ATLAS ATLAS TDAQTDAQ upgrade upgrade proposalproposal
TDAQ week at CERN
Michael Huffer, [email protected]
November 19, 2008
Department of Particle Physics & Astrophysics
2
OutlineOutline• DAQ support for next generation HEP experiments…
– “survey the requirements and capture their commonality”• One size does not fit all…
– generic building blocks• the (Reconfigurable) Cluster Element (RCE)• the Cluster Interconnect (CI)
– industry standard packaging• ATCA
• Packaged solutions– the RCE board– the CI board
• Applicability to ATLAS (the proposal):– motivation– scope– details
• ROM (Read-Out-Module)• CIM (Cluster-Interconnect-Module)• ROC (Read-Out-Crate)
– physical footprint, scaling & performance• Summary
Department of Particle Physics & Astrophysics
3
Three building block conceptsThree building block concepts
• Computational elements– must be low-cost
• $$$• footprint• power
– must support a variety of computational models
– must have both flexible and performanent I/O
• Mechanism to connect together these elements– must be low-cost– must provide low-latency/high-bandwidth
I/O – must be based on a commodity (industry)
protocol– must support a variety of interconnect
topologies• hierarchical• peer-to-peer• fan-In & fan-Out
• Packaging solution for both element & interconnect– must provide High Availability– must allow scaling– must support different physical I/O
interfaces– preferably based on a commercial standard
• The Reconfigurable Cluster Element (RCE)– employs System-On-Chip
technology (SOC)
• The Cluster Interconnect (CI)– based on 10-GE Ethernet
switching • ATCA
– Advanced Telecommunication Computing Architecture
– crate based, serial backplane
Department of Particle Physics & Astrophysics
4
(Reconfigurable) Cluster Element (RCE) (Reconfigurable) Cluster Element (RCE)
• Bundled software:– bootstrap loader– Open Source kernel
(RTEMS)• POSIX compliant
interfaces• standard I/P network
stack
– exception handling support
MGTs
Core
DSP tilesCombinatoric logic
Resources
Processor450 MHZ PPC-405
512 MByte RLD-II
Boot Options Memory Subsystem Configuration
data
128 MByte Flash
Data Exchange Interface (DEI)
instruction
reset & bootstrap
options
• Class libraries (C++) provide:– DEI support– configuration
interface
• Bundled software:– GNU cross-
development environment (C & C++)
– remote (network) GDB debugger
– network console
Department of Particle Physics & Astrophysics
5
ResourcesResources
• Multi-Gigabit Transceivers (MGTs)– up to 24 channels of:
• SER/DES• input/output buffering• clock recovery• 8b/10b encoder/decoder• 64b/66b encoder/decoder
– each channel can operate up to 6.5 gb/s– channels may be bound together for greater
aggregate speed• Combinatoric logic
• gates• flip-flops (block RAM)• I/O pins
• DSP support– contains up 192 Multiple-Accumulate-Add (MAC) units
Department of Particle Physics & Astrophysics
6
Derived configuration - Cluster Element (CE) Derived configuration - Cluster Element (CE)
Combinatoric logic
MGTs
Core
1.0/2/5/10.0 gb/s
PGP PGPPGP PGPPGPPGP PGPPGP
Ethernet MAC Ethernet MAC
MGTs
Combinatoric logic
E0
3.125 gb/s
E1
Department of Particle Physics & Astrophysics
7
The Cluster Interconnect (CI)The Cluster Interconnect (CI)
• Based on two Fulcrum FM224s– 24 port 10-GE switch– is an ASIC (packaging in 1433-ball BGA)– XAUI interface (supports multiple speeds including 100-
BaseT, 1-GE & 2.5 gb/s)– less then 24 watts at full capacity– cut-through architecture (packet ingress/egress < 200
NS)– full Layer-2 functionality (VLAN, multiple spanning tree
etc..)– configuration can be managed or unmanaged
Management bus
RCE
10-GE L2 switch10-GE L2 switch 10-GE L2 switch
Q0 Q1
Q2 Q3
Department of Particle Physics & Astrophysics
8
A A clustercluster of 12 elements of 12 elements
To back-end systems To back-end systems
ClusterInterconnect
Elements
From Front-End systems
switching fabric
switching fabric
Department of Particle Physics & Astrophysics
9
Why ATCA as a packaging standard?Why ATCA as a packaging standard?
• An emerging telecom standard… • Its attractive features:
– backplane & packaging available as a commercial solution– generous form factor
• 8U x 1.2” pitch– hot swap capability– well-defined environmental monitoring & control– emphasis on High Availability– external power input is low voltage DC
• allows for rack aggregation of power
• Its very attractive features:– the concept of a Rear Transition Module (RTM)
• allows all cabling to be on rear (module removal without interruption of cable plant)
• allows separation of data interface from the mechanism used to process that data
– high speed serial backplane• protocol agnostic• provision for different interconnect topologies
Department of Particle Physics & Astrophysics
10
RCE board + RTM (Block diagram)RCE board + RTM (Block diagram)
MFD RCE
flashmemory
RCE
slice0
slice1
slice2
slice3
slice7
slice6
slice5
slice4
MFD
Fiber-optic transceiversP2
Payload RTM
P3
P3
E1
E1
E0
E0
base
fabric
Department of Particle Physics & Astrophysics
11
RCE board + RTMRCE board + RTM
Media Carrier with flash
Media Slice controller
RCE
Zone 1(power)
Zone 2
Zone 3
transceivers
RTM
Department of Particle Physics & Astrophysics
12
Cluster Interconnect board + RTM (Block diagram)Cluster Interconnect board + RTM (Block diagram)
MFD CI
Q2
P2
1-GE
10-GE XFP
XFP
Q0
Q1 Q3
XFP
XFP
XFP
XFP
XFP
XFP
XFP
Payload RTM
P3
P3
10-GE1-GE
10-GE XFP
10-GE
(fabric)
(base)
base
fabric
(fabric)
(base)
Department of Particle Physics & Astrophysics
13
Cluster Interconnect board + RTMCluster Interconnect board + RTM
CI
Zone 3
Zone 1
10 GE switch
XFP
1G Ethernet
RCE
XFPRTM
Department of Particle Physics & Astrophysics
14
Typical (5 slot) ATCA crateTypical (5 slot) ATCA crate
fans
CI RTM
RCE RTM
CI board
Power suppliesRCE board
Shelf manager
Front
Back
Department of Particle Physics & Astrophysics
15
MotivationMotivation• Start with the premise that ROD replacement is
inevitable…– detector volume will scale upwards with luminosity– modularity of Front-End-Electronics will change
• Replacement is an opportunity to address additional concerns…– longevity of existing RODS
• long-term maintenance & support– many different ROD flavors
• difficult to capture commonality & reduce duplicated effort– ROS Ingress/Egress imbalance
• capable of almost 2 Gbytes/sec input• capable of less than 400/800 Mbytes/sec output
– scalability• each added ROD requires (on average) adding one ROS/PC
– one ROD (on average) drives two ROLs– one ROS/PC can process (roughly) 2 ROLs worth of data
• physical separation adds mechanical & operational constraints
Department of Particle Physics & Astrophysics
16
Scope & the Integrated Read-Out System Scope & the Integrated Read-Out System (IROS)(IROS)
• ROD crates– RODs– crate controller– L1 distribution & control– “back-plane” boards
• ROS/PC racks– ROS/PCs– ROBins
• ROLs (between ROS & ROD)• “wires” connecting these
components
• Proposal calls out for the replacement of the IROS…– Intrinsic modularity of the scheme allows replacing a subset
Level-2 (L2) trigger Event Builder (EB)
Integrated Read Out System (IROS)
Detector Front-End Electronics
• Upstream & downstream systems would remain the same…• Proposal is constructed out of three elements:
– ROM (Read-Out-Module)• combines functionality of ROD + ROS/PC
– CIM (Cluster-Interconnect-Module)– ROC (Read-Out-Crate)
Department of Particle Physics & Astrophysics
17
Read-Out-Module (ROM)Read-Out-Module (ROM)
From detector FEE
ROC backplane
Cluster Elements
P3
ROM
Rear Transition Module
P2
10-GE switch
10-GE switch
L1 fanout switch management
(X4) 3.2 gb/s
10 gb/s
(X2) 10 gb/s
(X2) 2.5 gb/s
CIM
Department of Particle Physics & Astrophysics
18
Read-Out-Crate (ROC)Read-Out-Crate (ROC)
from L1
CIM
Rear Transition Module
10-GE switch
P3
Backplane
Rear Transition Module
switch managementL1 fanout 10-GE
switch
Shelf Management
10-GE switch
10-GE switch
P3
ROMs
CIM
To monitoring & control from L1
To L2 & Event Building
switch managementL1 fanout
(X12) 10 gb/s
(X2) 10 gb/s
(X2) 2.5 gb/s
Department of Particle Physics & Astrophysics
19
IROSIROS
USA15
UX15
SDX1
IROS
Switching fabric Switching fabric
L1 trigger
Event Builder farm L2 farm
(X384) 3.2 gb/s
(X24) 10 gb/s
TileCalLAr Pixel SCT TRT MDT CSC
Inner Tracker MuonCalorimeter
Cal RPC TGC TPI CTP
Department of Particle Physics & Astrophysics
20
IROS plant footprintIROS plant footprint
Detector System
Detector Subsystem
Total RODs
Total ROLs
RODcrates
The ROS
Total ROMs
Total crates
Total CIMs
ROBIns PCs
Calorimeter
LArTileCal
192 762 16 255 68 48 4 8
32 64 4 24 8 8 1 2
Inner Detector
PixelSCTTRT
132 132 9 45 12 33 3 6
92 92 8 32 8 23 2 4
96 192 10 66 18 48 4 8
MuonMDTCSC
204 204 16 70 16 51 5 10
16 16 2 6 2 4 1 1
L1
CalorimeterMuon RPCMuon TGCMUCTPICTP
14 48 1 17 5 4 1 1
32 32 4 12 4 8 1 2
24 24 4 8 2 6 1 1
1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1
Totals 8361574
76 537145
235 25 45current proposed
Department of Particle Physics & Astrophysics
21
Scaling & performanceScaling & performance
Phase
Input event rates (KHz)Data size (Mbytes)
Output rates (Gbytes/s)
IROS L2 EB Event ROI L2 EB
0 75 75 2 1.5 .0162 1.2 3
1 100 100 3 3.0 .0826 8.3 9
2 100 100 6 5.8 .4884 48.8 35• Proposal scales linearly with number of ROMs…
– 2 Gbytes/sec/ROM for L2 network– .5 Gbytes/sec/ROM for Event Building network
• For plug replacement example this implies an output capacity of…– 270 Gbytes/sec for L2– 118 Gbytes/sec for Event Building
• As a comparison current system has a total output capacity of…– 116 Gbytes/sec (8 NIC channels)
• Performance requirements as a function of luminosity upgrade phase– numbers derived from Andy’s upgrade talk (TDAQ week,
May 2008)– both ROI size & number change as a function of luminosity
Department of Particle Physics & Astrophysics
22
SummarySummary• SLAC is positioning itself for a new generation of DAQ…
– strategy is based on the idea of modular building blocks• inexpensive computational elements (the RCE)• interconnect mechanism (the CI)• industry standard packaging (ATCA)
– architecture is now relatively mature• both demo boards (& corresponding RTMs) are functional • RTEMS ported & operating• network stack fully tested and functional
– performance and scaling meet expectations– costs have been established (engineering scales):
• ~$1K/RCE (goal is less then $750)• ~$1K/CI (goal is less then $750)
• This is an outside view looking in (presumptuous + sometimes useful)– Initiate discussion on the bigger picture– Separate proposal abstraction from its implementation
• common substrate ROD• integration of ROD + ROL + ROBin functionality
• Inherent modularity of this scheme allows piece-meal (adiabatic) replacement– can co-exist with current system
• Leverage recent industry innovation– System-On-chip (SOC)– High speed serial transmission– Low cost, small footprint, high-speed switching (10-GE)– Packaging standardization (serial backplanes and RTM)