1、INTERNATIONAL STANDARD ISO/IEC 13961 IEEE Std 1596 First edition 2000-07 Information technology Scalable Coherent Interface (SCI) Reference number ISO/IEC 13961:2000(E) IEEE Std 1596, 1998 EditionAbstract: The scalable coherent interface (SCI) provides computer-bus-like services but, instead of a bu
2、s, uses a collection of fast point-to-point unidirectional links to provide the far higher throughput needed for high-performance multiprocessor systems. SCI supports distributed, shared memory with optional cache coherence for tightly coupled systems, and message-passing for loosely coupled systems
3、. Initial SCI links are defined at 1 Gbyte/s (16-bit parallel) and 1 Gb/s (serial). For applications requiring modular packaging, an interchangeable module is specified along with connector and power. The packets and protocols that implement transactions are defined and their formal specification is
4、 provided in the form of computer programs. In addition to the usual read-and-write transactions, SCI supports efficient multiprocessor lock transactions. The distributed cache-coherence protocols are efficient and can recover from an arbitrary number of transmission failures. SCI protocols ensure f
5、orward progress despite multiprocessor conflicts (no deadlocks or starvation). Keywords: bus architecture, bus standard, cache coherence, distributed memory, fiber optic, interconnect, I/O system, link, mesh, multiprocessor, network, packet protocol, ring, seamless distributed computer, shared memor
6、y, switch, transaction set. The Institute of Electrical and Electronics Engineers, Inc. 345 East 47th Street, New York, NY 10017-2394, USA Copyright 1998 by the Institute of Electrical and Electronics Engineers, Inc. All rights reserved. First published in 1998. ISBN 2-8318-5375-3 No part of this pu
7、blication may be reproduced in any form, in an electronic retrieval system or otherwise, without the prior written permission of the publisher.INTERNATIONAL STANDARD ISO/IEC 13961 IEEE Std 1596 First edition 2000-07 Information technology Scalable Coherent Interface (SCI) Sponsor Microprocessor and
8、Microcomputer Standards Subcommittee of the IEEE Computer Society PRICE CODE XF For price, see current catalogue 2 ISO/IEC 13961:2000(E) IEEE Std 1596, 1998 Edition Copyright 1998 IEEE. All rights reserved. CONTENTS Page FOREWORD 12 Clause 1 Introduction 19 1.1 Document structure. 19 1.2 SCI overvie
9、w 20 1.2.1 Scope and directions. 20 1.2.2 The SCI approach. 21 1.2.3 System configurations. 22 1.2.4 Initial physical models. 23 1.2.5 SCI node model 24 1.2.6 Architectural parameters . 25 1.2.7 A common CSR architecture . 25 1.2.8 Structure of the specification. 26 1.3 Interconnect topologies. 26 1
10、.3.1 Bridged systems . 26 1.3.2 Scalable systems 27 1.3.3 Interconnected systems 27 1.3.4 Backplane rings 27 1.3.5 Interconnected rings . 28 1.3.6 Rectangular grid interconnects 29 1.3.7 Butterfly switches 30 1.3.8 Vendor-dependent switches 31 1.4 Transactions. 31 1.4.1 Packet formats 32 1.4.2 Input
11、 and output queues . 33 1.4.3 Request and response queues 34 1.4.4 Switch queues. 36 1.4.5 Subactions 36 1.4.6 Remote transactions (through agents). 39 1.4.7 Move transactions. 41 1.4.8 Broadcast moves 42 1.4.9 Broadcast passing by agents 43 1.4.10 Transaction types 44 1.4.11 Message passing 45 1.4.
12、12 Global clocks 45 1.4.13 Allocation protocols. 46 1.4.14 Queue allocation. 47 1.5 Cache coherence 49 1.5.1 Interconnect constraints 49 1.5.2 Distributed directories . 49 1.5.3 Standard optimizations 50 1.5.4 Future extensions . 50 1.5.5 TLB purges . 53ISO/IEC 13961:2000(E) 3 IEEE Std 1596, 1998 Ed
13、ition Copyright 1998 IEEE. All rights reserved. Clause Page 1.6 Reliability, availability, and support (RAS). 54 1.6.1 RAS overview . 54 1.6.2 Autoconfiguration 54 1.6.3 Control and status registers 54 1.6.4 Transmission-error detection and isolation 55 1.6.5 Error containment . 55 1.6.6 Hardware fa
14、ult retry (ringlet-local, physical layer option) . 56 1.6.7 Software fault recovery (end-to-end) . 56 1.6.8 System debugging 57 1.6.9 Alternate routing . 57 1.6.10 Online replacement. 57 2 References, glossary, and notation 58 2.1 Normative references 58 2.2 Conformance levels 58 2.3 Terms and defin
15、itions . 59 2.4 Bit and byte ordering. 66 2.5 Numerical values 68 2.6 C code 68 3 Logical protocols and formats 68 3.1 Packet formats 68 3.1.1 Packet types . 68 3.2 Send and echo packet formats 69 3.2.1 Request-send packet format . 69 3.2.2 Request-echo packet format . 72 3.2.3 Response-send packet
16、74 3.2.4 Standard status codes 76 3.2.5 Response-echo packet format. 78 3.2.6 Interconnect-affected fields. 79 3.2.7 Init packets . 80 3.2.8 Cyclic redundancy code (CRC) . 81 3.2.9 Parallel 16-bit CRC calculations 82 3.2.10 CRC stomping. 84 3.2.11 Idle symbols 85 3.3 Logical packet encodings 86 3.3.
17、1 Flag coding. 86 3.4 Transaction types . 89 3.4.1 Transaction commands . 89 3.4.2 Lock subcommands 92 3.4.3 Unaligned DMA transfers 94 3.4.4 Aligned block-transfer hints. 95 3.4.5 Move transactions. 97 3.4.6 Global time synchronization 98 3.5 Elastic buffers. 99 3.5.1 Elasticity models. 99 3.5.2 Id
18、le-symbol insertions . 100 3.5.3 Idle-symbol deletions 101 4 ISO/IEC 13961:2000(E) IEEE Std 1596, 1998 Edition Copyright 1998 IEEE. All rights reserved. Clause Page 3.6 Bandwidth allocation. 101 3.6.1 Fair bandwidth allocation 102 3.6.2 Setting ringlet priority 104 3.6.3 Bandwidth partitioning. 106
19、3.6.4 Types of transmission protocols 108 3.6.5 Pass-transmission protocol . 108 3.6.6 Low-transmission protocol. 111 3.6.7 Idle insertions . 114 3.6.8 High-transmission protocol 114 3.7 Queue allocation. 116 3.7.1 Queue reservations. 116 3.7.2 Multiple active sends. 118 3.7.3 Unfair reservations 11
20、9 3.7.4 Queue-selection protocols. 119 3.7.5 Re-send priorities 119 3.8 Transaction errors. 120 3.8.1 Requester timeouts (response-expected packets) . 120 3.8.2 Time-of-death timeout (optional, all nodes) . 120 3.8.3 Responder-processing errors 122 3.9 Transmission errors 123 3.9.1 Error isolation .
21、 123 3.9.2 Scrubber maintenance 125 3.9.3 Producer-detected errors 126 3.9.4 Consumer-detected errors. 128 3.10 Address initialization. 129 3.10.1 Transaction addressing. 129 3.10.2 Reset types. 131 3.10.3 Unique node identifiers . 132 3.10.4 Ringlet initialization. 133 3.10.5 Simple-subset ringlet
22、resets 135 3.10.6 Ringlet resets 135 3.10.7 Ringlet clears (optional) 137 3.10.8 Inserting initialization packets . 138 3.10.9 Address initialization. 139 3.11 Packet encoding . 140 3.11.1 Common encoding features (L18) . 140 3.11.2 Parallel encoding with 18 signals (P18). 141 3.11.3 Serial encoding
23、 with 20-bit symbols (S20). 141 3.12 SCI-specific control and status registers . 144 3.12.1 SCI transaction sets 144 3.12.2 SCI resets. 145 3.12.3 SCI-dependent fields within standard CSRs 145 3.12.4 SCI-dependent CSRs 148 3.12.5 SCI-dependent ROM. 151 3.12.6 Interrupt register formats. 155 3.12.7 I
24、nterleaved logical addressing. 157ISO/IEC 13961:2000(E) 5 IEEE Std 1596, 1998 Edition Copyright 1998 IEEE. All rights reserved. Clause Page 4 Cache-coherence protocols 158 4.1 Introduction. 158 4.1.1 Objectives. 158 4.1.2 SCI transaction components . 158 4.1.3 Physical addressing 159 4.1.4 Coherence
25、 directory overview. 159 4.1.5 Memory and cache tags 160 4.1.6 Instruction-execution model 161 4.1.7 Coherence document structure . 162 4.2 Coherence update sequences. 163 4.2.1 List prepend 163 4.2.2 List-entry deletion . 165 4.2.3 Update actions 167 4.2.4 Cache-line locks . 167 4.2.5 Stable sharin
26、g lists 168 4.3 Minimal-set coherence protocols. 171 4.3.1 Sharing-list updates 171 4.3.2 Cache fetching 171 4.3.3 Cache rollouts. 173 4.3.4 Instruction-execution model 174 4.4 Typical-set coherence protocols 175 4.4.1 Sharing-list updates 175 4.4.2 Read-only fetch. 175 4.4.3 Read-write fetch 177 4.
27、4.4 Data modifications 178 4.4.5 Mid and head deletions . 179 4.4.6 DMA reads and writes. 181 4.4.7 Instruction-execution model 183 4.5 Full-set coherence protocols. 184 4.5.1 Full-set option summary 184 4.5.2 CLEAN-list creation. 184 4.5.3 Sharing-list additions 185 4.5.4 Cache washing . 187 4.5.5
28、Cache flushing 189 4.5.6 Cache cleansing . 191 4.5.7 Pairwise sharing . 192 4.5.8 Pairwise-sharing faults 196 4.5.9 QOLB sharing . 197 4.5.10 Cache-access properties. 200 4.5.11 Instruction-execution model 201 4.6 C-code naming conventions 202 4.7 Coherent read and write transactions 203 4.7.1 Exten
29、ded mread transactions 204 4.7.2 Cache cread and cwrite64 transactions. 205 4.7.3 Smaller tag sizes 206 6 ISO/IEC 13961:2000(E) IEEE Std 1596, 1998 Edition Copyright 1998 IEEE. All rights reserved. Clause Page 5 C-code structure 207 5.1 Node structure 207 5.1.1 Signals within a node 207 5.1.2 Packet
30、 transfers among node components 208 5.1.3 Transfer-cloud components. 208 5.2 A nodes linc component . 210 5.2.1 A lincs subcomponents. 210 5.2.2 A lincs elastic buffer. 212 5.2.3 Other linc components 213 5.3 Other node components 213 5.3.1 A nodes core component 213 5.3.2 A nodes memory component
31、213 5.3.3 A nodes exec component . 214 5.3.4 A nodes proc component 215 6 Physical layers. 216 6.1 Type 1 module 217 6.1.1 Module characteristics 217 6.1.2 Module compatibility considerations 217 6.1.3 Module size. 218 6.1.4 Warpage, bowing, and deflection 224 6.1.5 Cooling . 225 6.1.6 Connector . 2
32、26 6.1.7 Power and ground connection. 227 6.1.8 Pin allocation for backplane parallel 18-signal encoding 229 6.1.9 Slot-identification signals 231 6.2 Type 18-DE-500 signals and power control. 232 6.2.1 SCI differential signals 233 6.2.2 Status lines. 233 6.2.3 Serial Bus signals . 233 6.2.4 Signal
33、levels and skew 233 6.2.5 Power-conversion control 236 6.3 Type 18-DE-500 module extender cable . 238 6.4 Type 18-DE-500 cable-link 240 6.5 Serial interconnection . 242 6.5.1 Serial interface Type 1-SE-1250, single-ended electrical. 243 6.5.2 Optical interface, fiber-optic signal type 1-FO-1250. 249
34、 6.5.3 Test methods 252 Annex A (informative) Ringlet initialization . 254 Annex B (informative) SCI design models. 257 B.1 Fast counters. 257 B.2 Translation-lookaside-buffer coherence . 257 B.3 Coherent lock models 261 B.4 Coherence-performance models 263 Bibliography . 265ISO/IEC 13961:2000(E) 7
35、IEEE Std 1596, 1998 Edition Copyright 1998 IEEE. All rights reserved. Page Figure 1 Physical-layer alternatives. 23 Figure 2 SCI node model. 24 Figure 3 64-bit-fixed addressing 25 Figure 4 Bridged systems 26 Figure 5 Backplane rings. 28 Figure 6 Interconnected rings 29 Figure 7 2-D processor grids .
36、 29 Figure 8 Butterfly ringlets 30 Figure 9 Switch interface. 31 Figure 10 Subactions. 32 Figure 11 Send-packet format, simplified. 32 Figure 12 Responder queues. 34 Figure 13 Logical requester/responder queues 35 Figure 14 Paired request and response queues. 35 Figure 15 Basic SCI bridge, paired re
37、quest and response queues . 36 Figure 16 Local transaction components . 37 Figure 17 Local transaction components (busied by responder). 38 Figure 18 Remote transaction components 40 Figure 19 Remote move-transaction components 41 Figure 20 Broadcast starts 43 Figure 21 Broadcast resumes 43 Figure 2
38、2 Transaction formats. 44 Figure 23 Bandwidth partitioning . 46 Figure 24 Resource bottlenecks 47 Figure 25 Queue allocation avoids starvation 48 Figure 26 Distributed cache tags . 49 Figure 27 Request combining 52 Figure 28 Binary tree. 52 Figure 29 TLB purging. 53 Figure 30 Hardware fault-retry se
39、quence. 56 Figure 31 Software fault-retry on coherent data. 57 Figure 32 Big-endian packet notation 67 Figure 33 Big-endian register notation. 67 Figure 34 Send- and echo-packet formats . 69 Figure 35 Request-packet format. 70 Figure 36 Request-packet symbols 70 Figure 37 Request-echo packet format
40、72 Figure 38 Response-packet format 74 Figure 39 Response-packet symbols . 75 Figure 40 Response-echo packet format . 78 Figure 41 Initialization-packet format. 80 Figure 42 Initialization-packet format example (companyId-based uniqueId value) 81 Figure 43 Serialized implementation of 16-bit CRC 82
41、Figure 44 Parallel CRC check 84 Figure 45 Remote transaction components (local request-send damaged). 85 Figure 46 Logical idle-symbol encoding. 85 Figure 47 Flag framing convention. 86 Figure 48 Logical send- and init-packet framing convention. 87 Figure 49 Logical echo-packet framing convention 87
42、 Figure 50 Logical sync-packet framing convention. 88 Figure 51 Logical abort-packet framing convention 88 Figure 52 Selected-byte reads and writes 91 Figure 53 Simplified lock model. 92 Figure 54 Selected-byte locks (quadlet access). 93 Figure 55 Selected-byte locks (octlet access) 94 Figure 56 Exp
43、ected DMA read transfers 94 Figure 57 Expected DMA write transfers 95 Figure 58 DMA block-transfer model 96 Figure 59 Time-sync on SCI 98 Figure 60 Elasticity model . 99 8 ISO/IEC 13961:2000(E) IEEE Std 1596, 1998 Edition Copyright 1998 IEEE. All rights reserved. Page Figure 61 Input-synchronizer mo
44、del. 100 Figure 62 Idle-symbol insertion 100 Figure 63 Idle-symbol deletion. 101 Figure 64 Fair bandwidth allocation. 103 Figure 65 Increasing ringlet priority . 105 Figure 66 Restoring ringlet priority. 105 Figure 67 Idle-symbol creation, fair-only node . 106 Figure 68 Idle-symbol creation, unfair-
45、capable node 107 Figure 69 Idle consumption, fair-only node 107 Figure 70 Idle consumption, unfair-capable node. 108 Figure 71 Pass-transmission model (fair-only node) 109 Figure 72 Pass-transmission enabled 109 Figure 73 Pass-transmission active . 110 Figure 74 Pass-transmission recovery. 110 Figur
46、e 75 Low/high-transmission model. 111 Figure 76 Low-transmission enabled . 111 Figure 77 Low-transmission active. 112 Figure 78 Low/high-transmission recovery. 113 Figure 79 Low/high-transmission debt repayment 113 Figure 80 Low/high-transmission idle insertion 114 Figure 81 High-transmission enable
47、d. 115 Figure 82 Consumer send-packet queue reservations . 116 Figure 83 A/B age labels . 118 Figure 84 Response timeouts (request and no response) 120 Figure 85 Time-of-death discards 121 Figure 85 Packet life-cycle intervals 121 Figure 87 Time-of-death generation model 122 Figure 88 Responders add
48、ress-error processing. 122 Figure 89 Response timeouts (request and no response) 123 Figure 90 Error-logging registers . 124 Figure 91 Scrubber maintenance functions 125 Figure 92 Detecting lost low-go bits. 126 Figure 93 Producers address-error processing . 127 Figure 94 Producers echo-timeout proc
49、essing 127 Figure 95 Producer fatal-error recovery (optional) . 128 Figure 96 Consumer error recovery . 129 Figure 97 SCI (64-bit fixed) addressing . 129 Figure 98 Forms of node resets. 132 Figure 99 Receiver synchronization and scrubber selection. 134 Figure 100 Reset-closure generates idle symbols 134 Figure 100 Idle-closure injec
