Files
gem5/configs/dram/sweep.py
Wendy Elsasser d228a283c9 mem: Modify DRAM controller for flexibility and new memories
This change includes:
1) Verify available command bandwidth
2) Add support for multi-cycle commands
3) Add new timing parameters
4) Add ability to interleave bursts
5) Add LPDDR5 configurations

The DRAM controller historically does not verify contention on the
command bus and if there is adaquate command bandwidth to issue a
new command. As memory technologies evolve, multiple cycles are becoming
a requirement for some commands.  Depending on the burst length, this
can stress the command bandwidth. A check was added to verify command
issue does not exceed a maximum value within a defined window. The
default window is a burst, with the maximum value defined based on the
burst length and media clocking characteristics. When the command bandwidth
is exceeded, commands will be shifted to subsequent burst windows.

Added support for multi-cycle commands, specifically Activate, which
requires a larger address width as capacities grow.  Additionally,
added support for multi-cycle Read / Write bursts for low power
DRAM cases in which additional CLK synchronization may be required
to run at higher speeds.

To support emerging memories, added the following new timing parameters.
1) tPPD -- Precharge-to-Precharge delay
2) tAAD -- Max delay between Activate-1 and Activate-2 commands

I/O data rates are continuing to increase for DRAM but the core frequency
is still fairly stagnant for many technologies. As we increase the burst
length, either the core prefetch needs to increase (for a seamless burst)
or the burst will be transferred with gaps on the data bus. To support
the latter case, added the ability to interleave 2 bursts across bank
groups.

Using the changes above, added an initial set of LPDDR5 configurations.

Change-Id: I1b14fed221350e6e403f7cbf089fe6c7f033c181
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/26236
Reviewed-by: Matthew Poremba <matthew.poremba@amd.com>
Reviewed-by: Jason Lowe-Power <power.jg@gmail.com>
Maintainer: Jason Lowe-Power <power.jg@gmail.com>
Tested-by: Gem5 Cloud Project GCB service account <345032938727@cloudbuild.gserviceaccount.com>
Tested-by: kokoro <noreply+kokoro@google.com>
2020-04-09 16:15:07 +00:00

200 lines
7.6 KiB
Python

# Copyright (c) 2014-2015, 2018-2020 ARM Limited
# All rights reserved.
#
# The license below extends only to copyright in the software and shall
# not be construed as granting a license to any other intellectual
# property including but not limited to intellectual property relating
# to a hardware implementation of the functionality of the software
# licensed hereunder. You may use the software subject to the license
# terms below provided that you ensure that this notice is replicated
# unmodified and in its entirety in all distributions of the software,
# modified or unmodified, in source code or in binary form.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are
# met: redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer;
# redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution;
# neither the name of the copyright holders nor the names of its
# contributors may be used to endorse or promote products derived from
# this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
from __future__ import print_function
from __future__ import absolute_import
import math
import optparse
import m5
from m5.objects import *
from m5.util import addToPath
from m5.stats import periodicStatDump
addToPath('../')
from common import ObjectList
from common import MemConfig
# this script is helpful to sweep the efficiency of a specific memory
# controller configuration, by varying the number of banks accessed,
# and the sequential stride size (how many bytes per activate), and
# observe what bus utilisation (bandwidth) is achieved
parser = optparse.OptionParser()
dram_generators = {
"DRAM" : lambda x: x.createDram,
"DRAM_ROTATE" : lambda x: x.createDramRot,
}
# Use a single-channel DDR3-1600 x64 (8x8 topology) by default
parser.add_option("--mem-type", type="choice", default="DDR3_1600_8x8",
choices=ObjectList.mem_list.get_names(),
help = "type of memory to use")
parser.add_option("--mem-ranks", "-r", type="int", default=1,
help = "Number of ranks to iterate across")
parser.add_option("--rd_perc", type="int", default=100,
help = "Percentage of read commands")
parser.add_option("--mode", type="choice", default="DRAM",
choices=dram_generators.keys(),
help = "DRAM: Random traffic; \
DRAM_ROTATE: Traffic rotating across banks and ranks")
parser.add_option("--addr-map", type="choice",
choices=ObjectList.dram_addr_map_list.get_names(),
default="RoRaBaCoCh", help = "DRAM address map policy")
(options, args) = parser.parse_args()
if args:
print("Error: script doesn't take any positional arguments")
sys.exit(1)
# at the moment we stay with the default open-adaptive page policy,
# and address mapping
# start with the system itself, using a multi-layer 2.0 GHz
# crossbar, delivering 64 bytes / 3 cycles (one header cycle)
# which amounts to 42.7 GByte/s per layer and thus per port
system = System(membus = IOXBar(width = 32))
system.clk_domain = SrcClockDomain(clock = '2.0GHz',
voltage_domain =
VoltageDomain(voltage = '1V'))
# we are fine with 256 MB memory for now
mem_range = AddrRange('256MB')
system.mem_ranges = [mem_range]
# do not worry about reserving space for the backing store
system.mmap_using_noreserve = True
# force a single channel to match the assumptions in the DRAM traffic
# generator
options.mem_channels = 1
options.external_memory_system = 0
options.tlm_memory = 0
options.elastic_trace_en = 0
MemConfig.config_mem(options, system)
# the following assumes that we are using the native DRAM
# controller, check to be sure
if not isinstance(system.mem_ctrls[0], m5.objects.DRAMCtrl):
fatal("This script assumes the memory is a DRAMCtrl subclass")
# there is no point slowing things down by saving any data
system.mem_ctrls[0].null = True
# Set the address mapping based on input argument
system.mem_ctrls[0].addr_mapping = options.addr_map
# stay in each state for 0.25 ms, long enough to warm things up, and
# short enough to avoid hitting a refresh
period = 250000000
# stay in each state as long as the dump/reset period, use the entire
# range, issue transactions of the right DRAM burst size, and match
# the DRAM maximum bandwidth to ensure that it is saturated
# get the number of banks
nbr_banks = system.mem_ctrls[0].banks_per_rank.value
# determine the burst length in bytes
burst_size = int((system.mem_ctrls[0].devices_per_rank.value *
system.mem_ctrls[0].device_bus_width.value *
system.mem_ctrls[0].burst_length.value) / 8)
# next, get the page size in bytes
page_size = system.mem_ctrls[0].devices_per_rank.value * \
system.mem_ctrls[0].device_rowbuffer_size.value
# match the maximum bandwidth of the memory, the parameter is in seconds
# and we need it in ticks (ps)
itt = getattr(system.mem_ctrls[0].tBURST_MIN, 'value',
system.mem_ctrls[0].tBURST.value) * 1000000000000
# assume we start at 0
max_addr = mem_range.end
# use min of the page size and 512 bytes as that should be more than
# enough
max_stride = min(512, page_size)
# create a traffic generator, and point it to the file we just created
system.tgen = PyTrafficGen()
# add a communication monitor
system.monitor = CommMonitor()
# connect the traffic generator to the bus via a communication monitor
system.tgen.port = system.monitor.slave
system.monitor.master = system.membus.slave
# connect the system port even if it is not used in this example
system.system_port = system.membus.slave
# every period, dump and reset all stats
periodicStatDump(period)
# run Forrest, run!
root = Root(full_system = False, system = system)
root.system.mem_mode = 'timing'
m5.instantiate()
def trace():
addr_map = ObjectList.dram_addr_map_list.get(options.addr_map)
generator = dram_generators[options.mode](system.tgen)
for stride_size in range(burst_size, max_stride + 1, burst_size):
for bank in range(1, nbr_banks + 1):
num_seq_pkts = int(math.ceil(float(stride_size) / burst_size))
yield generator(period,
0, max_addr, burst_size, int(itt), int(itt),
options.rd_perc, 0,
num_seq_pkts, page_size, nbr_banks, bank,
addr_map, options.mem_ranks)
yield system.tgen.createExit(0)
system.tgen.start(trace())
m5.simulate()
print("DRAM sweep with burst: %d, banks: %d, max stride: %d, request \
generation period: %d" % (burst_size, nbr_banks, max_stride, itt))