smmzhang / sim-d Goto Github PK

Sim-D is a (GPU-like) wide-SIMD architecture tailored for hard real-time
systems. Inspired by the PRedictable Execution Model (PREM) [1], Sim-D enforces
isolation between compute- and storage resources, allowing them to be occupied
in parallel free of interference. By treating a work-group as a sequence of
compute- and access phases, the problem of WCET analysis becomes a problem of
scheduling the phases for each work-group in the kernel-instance. Guarantees on
DRAM request times are facilitated with explicitly-coalesced large DRAM
transfers, requesting the data (tiles or indirect) for an entire work-group at
a time. This permits using a closed-page DRAM policy on the boundaries of a
request, while still allowing the use of deterministic burst request
re-ordering policies within a request. As a result, a 4KiB tile of data can be
performed at a guaranteed DRAM bus utilisation of ~72%.

Publication of the accompanying research work is currently under review.

This repository contains both the cycle-accurate simulation model for Sim-D, as
well as tools that perform WCET analysis on kernels written for Sim-D. This
source code depends on three external libraries: SystemC, Ramulator[2] and
DRAMPower[3]. Please refer to the INSTALL file for more details on bulding/
installing Sim-D and its dependencies. Sim-D's source code is licensed under the
terms of the GPLv3 or later, of which a copy is included in the LICENSE file.

Usage
===

After build, two tools can be found in the root directory:
- main: Performs simulation of a kernel-instance,
- wcet: Performs WCET analysis of kernel-instance.

For a complete overview of the parameters for each tool, run them without
arguments. For convenience, the launch/sim and launch/wcet contain bash-scripts
that invoke these tools for a given benchmark with the correct NDRange and
work-group dimensions, as well as the parameters needed to validate the
kernel-instance's output. These scripts should be invoked from the Sim-D root
directory.

Several auxiliary tools will be built:
- src/mc/mc: simulate a single tiled DRAM request,
- src/mc/mcIdx: simulate an iterative indexed DRAM request,
- src/util/ddr4_lid: Compute the execution time of DRAM requests for a pattern-
  based DRAM controller [4],
- src/util/cmp_kfusion_track: Performs output buffer validation for the KFusion
  track kernel-instance output,
- src/isa/print: Generate LaTeX-formatted documentation about the ISA.
- src/stridegen/stridegen: Generate and print the burst requests resulting from
  a given stride requests, and estimate their longeset isue delay and energy
  cost when processed by a pattern-based DRAM controller
- src/stridegen/stridegen_sp: Generate and print the line requests for a given
  stride request.

To run unit-tests, build Sim-D with a "Debug" build type. This can easily be
configured using ccmake. After building, tests can be executed with "make test"
or "ninja test". We recommend configuring Sim-D's build type as either "Release"
or "RelWithDebugInfo" when intending to execute simulation or WCET analysis for
performance reasons.

References
===

[1] R. Pellizzoni, E. Betti, S. Bak, G. Yao, J. Criswell, M. Caccamo, and R.
Kegley, “A Predictable Execution Model for COTS-Based Embedded Systems,” in 17th
IEEE Real-Time and Embedded Technology and Applications Symp., April 2011
[2] Y. Kim, W. Yang, and O. Mutlu, “Ramulator: A fast and extensible dram
simulator,” IEEE Computer Architecture Letters, vol. 15, no. 1, Jan 2016.
[3] K. Chandrasekar, C. Weis, Y. Li, S. Goossens, M. Jung, O. Naji, B. Akesson,
N. Wehn, and K. Goossens. DRAMPower: Open-source DRAM Power & Energy Estimation
Tool, 2012.
[4] B. Akesson, K. Goossens, and M. Ringhofer, “Predator: A Predictable SDRAM
Memory Controller,” in Proc. 5th IEEE/ACM Int. Conf. on Hardware/Software
Codesign and System Synthesis. 2007