TBS Simulator

• TBS simulator aims to simulate GPGPU's thread block scheduling.
• This simuation currently supports 6 scheduling algorithms.
• A new algorithm can be easily added by implementing a few framework routines.
• Multiple instances of computing or memory resources are newly supported.

Build

• requires gcc, make build tools running on Linux OS
• # make

Run

• ./simtbs <options> <config path>
• options list
  • -h shows help messages
  • -p <tbs policy> : assign TBS policy
    available policies: rr(default), rrf, bfa, dfa
  • -v : verbose mode
  • -g <new conf path> : auto workload generation mode
• A template configuration file is provieded as simtbs.conf.tmpl
• ./simtbs sample.conf
  • run simtbs with rr policy using sample.conf configuration file
• ./simtbs -p bfa sample.conf
  • run simtbs with bfa policy
• ./simtbs -g newsim.conf sample.conf
  • automatically generate kernel workloads based on sample.conf
  • newsim.conf file is newly created

Configuration File Syntax

• comment line: starting with # at a first column
• section line: starting with * at a first column
  • supported sections: general, workload, sm, mem, overhead, kernel
  • section has one or more lines
  • each line has one or more fields separated by blank
• general section: general configuration
  • Currently, only a maximum simulation time is supported.
  • a maximum simulation time should be provided in auto workload generation mode.
• workload section: settings for automatic workload generation mode
  • format: level n_tbs_range tb_duration_range sm_req_spec rsc_req n_tbs_range tb_duration_range`
  • level: minimum SM usage generating workload (1~100). If only a level is provided, kernels are generated randomly from kernel sections.
  • n_tbs_range: requested TB range for kernel
  • tb_duration_range: single TB execution time range
  • sm_rsc_req: computing resource amount per each scheduled TB count. Multiple entries can be allowed. range(1-8) or multiple(1,3,8) format is possible.
• sm section: SM configuration
  • format: sm_count n_rscs_sched n_rscs_compute sm_rsc_max
  • sm_count: total SM count in a GPGPU device
  • n_rscs_sched: number of computing resources for scheduling.
  • n_rscs_compute: number of computing resources, which are scheduled beyond their max usages.
  • sm_rsc_max: maximum computing resource count per SM. multiple items can be listed if n_rscs_sched + n_rscs_compute > 1
• mem section: memory sizes for various memory types
  • gobal/shared/etc memory size. multiple items should be given if multiple memory resources are simulated.
• overhead section: define TB execution overhead depending on SM computing resource usage
  • this seciton must be provided
  • each line has a used resource count and overhead.
  • all lines should be sorted by a used resource count.
  • 3 0.1 means that 10% TB execution overhead is added if 3 or less resources in SM are used.
• kernel section: define kernels for simulation
  • each line has start timestamp, TB count in kernel, computing resource count and TB execution time(in timestamp)

Scheduling Policy

rr(Round Robin)

• assign TB to SM in round robin fashion

rrf(Round Robin Fully)

• same as rr but move to the next SM only if crrent SM is fully used

bfa(Breadth First Allocation)

• schedule TB to be spreaded to SM
• prioritize SM which has the smallest resource usage
• maintain balanced resource usage for all SM's

dfa(Depth First Allocation)

• schedule TB to be converged to the most used SM
• prioritize SM which has the largest resource usage

smk(Simultaneous Multi Kernel)

• Support multiple resources but assign TB to SM in round robin fashion

fua(Function-unit Aware)

• Support multiple resources
• May schedule the same computing-bound TB if fine-grained resources are availalbe at function unit level.