Difference between revisions of "TraceCPU"

From gem5
Jump to: navigation, search
(Created page with " == Headline text == Introduction The goal of this project is to achieve performance estimation in a fast and accurate way by capturing and replaying dependency and timing an...")
 
m (→‎Overview)
 
(12 intermediate revisions by one other user not shown)
Line 1: Line 1:
 +
__TOC__
 +
== Overview ==
  
−
== Headline text ==
+
The Trace CPU model plays back ''elastic traces'', which are dependency and timing annotated traces generated by the Elastic Trace Probe attached to the O3 CPU model. The focus of the Trace CPU model is to achieve memory-system (cache-hierarchy, interconnects and main memory) performance exploration in a fast and reasonably accurate way instead of using the detailed but slow O3 CPU model.
−
Introduction
+
The traces have been developed for single-threaded benchmarks simulating in both SE and FS mode. They have been correlated for 15 memory-sensitive SPEC 2006 benchmarks and a handful of HPC proxy apps by interfacing the Trace CPU with classic memory system and varying cache design parameters and DRAM memory type. In general, elastic traces can be ported to other simulation environments.
  
−
The goal of this project is to achieve performance estimation in a fast and accurate way by capturing and replaying dependency and timing annotated traces instead of using a detailed but slow CPU model. The focus is on data memory accesses which are presumably the bulk traffic to conduct performance estimation studies of different interconnect and memories.
+
'''Publication'''
 +
 
 +
[https://doi.org/10.1109/SAMOS.2016.7818336 "Exploring System Performance using Elastic Traces: Fast, Accurate and Portable"]
 +
Radhika Jagtap, Stephan Diestelhorst, Andreas Hansson, Matthias Jung and Norbert Wehn
 +
SAMOS 2016
 +
 
 +
'''Trace generation and replay methodology'''
 +
 
 +
[[File:Etrace_methodology.jpg|800px]]
 +
 
 +
== Elastic Trace Generation ==
 +
 
 +
The Elastic Trace Probe Listener listens to Probe Points inserted in O3 CPU pipeline stages. It monitors each instruction and creates a dependency graph by recording data Read-After-Write dependencies and order dependencies between loads and stores. It writes the instruction fetch request trace and the elastic data memory request trace as two separate files as shown below.
 +
 
 +
[[File:Etraces_output.jpg|800px]]
 +
 
 +
=== Trace file formats ===
 +
 
 +
The elastic data memory trace and fetch request trace are both encoded using google protobuf.
 +
 
 +
{|class="wikitable"
 +
|+Elastic Trace fields in protobuf format
 +
|required uint64 seq_num || Instruction number used as an id for tracking dependencies
 +
|-
 +
|required RecordType type || RecordType enum has values: INVALID, LOAD, STORE, COMP
 +
|-
 +
| optional uint64 p_addr  || Physical memory address if instruction is a load/store
 +
|-
 +
| optional uint32 size  || Size in bytes of data if instruction is a load/store
 +
|-
 +
| optional uint32 flags  || Flags or attributes of the access, ex. Uncacheable
 +
|-
 +
| required uint64 rob_dep  || Past instruction number on which there is order (ROB) dependency
 +
|-
 +
| required uint64 comp_delay  || Execution delay between the completion of the last dependency and the execution of the instruction
 +
|-
 +
| repeated uint64 reg_dep  || Past instruction number on which there is RAW data dependency
 +
|-
 +
| optional uint32 weight  || To account for committed instructions that were filtered out
 +
|-
 +
| optional uint64 pc || Instruction address, i.e. the program counter
 +
|-
 +
| optional uint64 v_addr || Virtual memory address if instruction is a load/store
 +
|-
 +
| optional uint32 asid || Address Space ID
 +
|}
 +
 
 +
 
 +
A decode script in Python is available at util/decode_inst_dep_trace.py that outputs the trace in ASCII format.
 +
 
 +
'''Example of a trace in ASCII'''
 +
{|
 +
|1,356521,COMP,8500::
 +
|-
 +
|2,35656,1,COMP,0:,1:
 +
|-
 +
|3,35660,1,LOAD,1748752,4,74,500:,2:
 +
|-
 +
|4,35660,1,COMP,0:,3:
 +
|-
 +
|5,35664,1,COMP,3000::,4
 +
|-
 +
|6,35666,1,STORE,1748752,4,74,1000:,3:,4,5
 +
|-
 +
|7,35666,1,COMP,3000::,4
 +
|-
 +
|8,35670,1,STORE,1748748,4,74,0:,6,3:,7
 +
|-
 +
|9,35670,1,COMP,500::,7
 +
|}
 +
 
 +
Each record in the instruction fetch trace has the following fields.
 +
 
 +
{|class="wikitable"
 +
|+Instruction fetch trace fields in protobuf format
 +
| required uint64 tick  || Timestamp of the access
 +
|-
 +
| required uint32 cmd  || Read or Write (in this case always Read)
 +
|-
 +
| required uint64 addr  || Physical memory address
 +
|-
 +
| required uint32 size  || Size in bytes of data
 +
|-
 +
| optional uint32 flags  || Flags or attributes of the access
 +
|-
 +
| optional uint64 pkt_id  || Id of the access
 +
|-
 +
| optional uint64 pc  || Instruction address, i.e. the program counter
 +
|}
 +
 
 +
The decode script in Python at util/decode_packet_trace.py can be used to output the trace in ASCII format.
 +
 
 +
'''Compile dependencies'''
 +
 
 +
You need to install google protocol buffer as the traces are recorded using this.
 +
 
 +
sudo apt-get install protobuf-compiler<br />sudo apt-get install libprotobuf-dev
 +
 
 +
=== Scripts and options ===
 +
 
 +
* SE mode
 +
** build/ARM/gem5.opt [gem5.opt options] -d bzip_10Minsts configs/example/se.py [se.py options] --cpu-type=arm_detailed --caches --cmd=$M5_PATH/binaries/arm_arm/linux/bzip2 --options=$M5_PATH/data/bzip2/lgred/input/input.source -I 10000000 --elastic-trace-en --data-trace-file=deptrace.proto.gz --inst-trace-file=fetchtrace.proto.gz --mem-type=SimpleMemory
 +
 
 +
* FS mode: Create a checkpoint for your region of interest and resume from the checkpoint but with O3 CPU model and tracing enabled
 +
** build/ARM/gem5.opt --outdir=m5out/bbench ./configs/example/fs.py [fs.py options] --benchmark bbench-ics
 +
** build/ARM/gem5.opt --outdir=m5out/bbench/capture_10M ./configs/example/fs.py [fs.py options] --cpu-type=arm_detailed --caches --elastic-trace-en --data-trace-file=deptrace.proto.gz --inst-trace-file=fetchtrace.proto.gz --mem-type=SimpleMemory --checkpoint-dir=m5out/bbench -r 0 --benchmark bbench-ics -I 10000000
 +
 
 +
== Replay with Trace CPU ==
 +
 
 +
The execution trace generated above is then consumed by the Trace CPU as illustrated below.
 +
 
 +
[[File:Trace_cpu_top_level.jpg|699px]]
 +
 
 +
The Trace CPU model inherits from the Base CPU and interfaces with data and instruction L1 caches. A diagram of the Trace CPU explaining the major logic and control blocks is shown below.
 +
 
 +
[[File:Trace_cpu_detail.jpg|800px]]
 +
 
 +
=== Scripts and options ===
 +
* A trace replay script in the examples folder can be used to play back SE and FS generated traces
 +
** build/ARM/gem5.opt [gem5.opt options] -d bzip_10Minsts_replay configs/example/etrace_replay.py [options] --cpu-type=trace --caches --data-trace-file=bzip_10Minsts/deptrace.proto.gz --inst-trace-file=bzip_10Minsts/fetchtrace.proto.gz --mem-size=4GB

Latest revision as of 16:45, 18 March 2017

Overview

The Trace CPU model plays back elastic traces, which are dependency and timing annotated traces generated by the Elastic Trace Probe attached to the O3 CPU model. The focus of the Trace CPU model is to achieve memory-system (cache-hierarchy, interconnects and main memory) performance exploration in a fast and reasonably accurate way instead of using the detailed but slow O3 CPU model. The traces have been developed for single-threaded benchmarks simulating in both SE and FS mode. They have been correlated for 15 memory-sensitive SPEC 2006 benchmarks and a handful of HPC proxy apps by interfacing the Trace CPU with classic memory system and varying cache design parameters and DRAM memory type. In general, elastic traces can be ported to other simulation environments.

Publication

"Exploring System Performance using Elastic Traces: Fast, Accurate and Portable" Radhika Jagtap, Stephan Diestelhorst, Andreas Hansson, Matthias Jung and Norbert Wehn SAMOS 2016

Trace generation and replay methodology

Etrace methodology.jpg

Elastic Trace Generation

The Elastic Trace Probe Listener listens to Probe Points inserted in O3 CPU pipeline stages. It monitors each instruction and creates a dependency graph by recording data Read-After-Write dependencies and order dependencies between loads and stores. It writes the instruction fetch request trace and the elastic data memory request trace as two separate files as shown below.

Etraces output.jpg

Trace file formats

The elastic data memory trace and fetch request trace are both encoded using google protobuf.

Elastic Trace fields in protobuf format
required uint64 seq_num Instruction number used as an id for tracking dependencies
required RecordType type RecordType enum has values: INVALID, LOAD, STORE, COMP
optional uint64 p_addr Physical memory address if instruction is a load/store
optional uint32 size Size in bytes of data if instruction is a load/store
optional uint32 flags Flags or attributes of the access, ex. Uncacheable
required uint64 rob_dep Past instruction number on which there is order (ROB) dependency
required uint64 comp_delay Execution delay between the completion of the last dependency and the execution of the instruction
repeated uint64 reg_dep Past instruction number on which there is RAW data dependency
optional uint32 weight To account for committed instructions that were filtered out
optional uint64 pc Instruction address, i.e. the program counter
optional uint64 v_addr Virtual memory address if instruction is a load/store
optional uint32 asid Address Space ID


A decode script in Python is available at util/decode_inst_dep_trace.py that outputs the trace in ASCII format.

Example of a trace in ASCII

1,356521,COMP,8500::
2,35656,1,COMP,0:,1:
3,35660,1,LOAD,1748752,4,74,500:,2:
4,35660,1,COMP,0:,3:
5,35664,1,COMP,3000::,4
6,35666,1,STORE,1748752,4,74,1000:,3:,4,5
7,35666,1,COMP,3000::,4
8,35670,1,STORE,1748748,4,74,0:,6,3:,7
9,35670,1,COMP,500::,7

Each record in the instruction fetch trace has the following fields.

Instruction fetch trace fields in protobuf format
required uint64 tick Timestamp of the access
required uint32 cmd Read or Write (in this case always Read)
required uint64 addr Physical memory address
required uint32 size Size in bytes of data
optional uint32 flags Flags or attributes of the access
optional uint64 pkt_id Id of the access
optional uint64 pc Instruction address, i.e. the program counter

The decode script in Python at util/decode_packet_trace.py can be used to output the trace in ASCII format.

Compile dependencies

You need to install google protocol buffer as the traces are recorded using this.

sudo apt-get install protobuf-compiler
sudo apt-get install libprotobuf-dev

Scripts and options

  • SE mode
    • build/ARM/gem5.opt [gem5.opt options] -d bzip_10Minsts configs/example/se.py [se.py options] --cpu-type=arm_detailed --caches --cmd=$M5_PATH/binaries/arm_arm/linux/bzip2 --options=$M5_PATH/data/bzip2/lgred/input/input.source -I 10000000 --elastic-trace-en --data-trace-file=deptrace.proto.gz --inst-trace-file=fetchtrace.proto.gz --mem-type=SimpleMemory
  • FS mode: Create a checkpoint for your region of interest and resume from the checkpoint but with O3 CPU model and tracing enabled
    • build/ARM/gem5.opt --outdir=m5out/bbench ./configs/example/fs.py [fs.py options] --benchmark bbench-ics
    • build/ARM/gem5.opt --outdir=m5out/bbench/capture_10M ./configs/example/fs.py [fs.py options] --cpu-type=arm_detailed --caches --elastic-trace-en --data-trace-file=deptrace.proto.gz --inst-trace-file=fetchtrace.proto.gz --mem-type=SimpleMemory --checkpoint-dir=m5out/bbench -r 0 --benchmark bbench-ics -I 10000000

Replay with Trace CPU

The execution trace generated above is then consumed by the Trace CPU as illustrated below.

Trace cpu top level.jpg

The Trace CPU model inherits from the Base CPU and interfaces with data and instruction L1 caches. A diagram of the Trace CPU explaining the major logic and control blocks is shown below.

Trace cpu detail.jpg

Scripts and options

  • A trace replay script in the examples folder can be used to play back SE and FS generated traces
    • build/ARM/gem5.opt [gem5.opt options] -d bzip_10Minsts_replay configs/example/etrace_replay.py [options] --cpu-type=trace --caches --data-trace-file=bzip_10Minsts/deptrace.proto.gz --inst-trace-file=bzip_10Minsts/fetchtrace.proto.gz --mem-size=4GB
Retrieved from "http://daystrom.m5sim.org/wiki/index.php?title=TraceCPU&oldid=8146"