Model Configuration Settings¶
The C++ only models (cpp_model/) offer pure functionality modeling.
An instruction "completes" the moment it starts, with no notion of how
long it would actually take on real hardware. The Sitar models
(sitar_model/) additionally offer component-wise latency modeling:
independent, configurable delay values for opcode execution, each
memory-access channel, and, where present, the MMU and its interconnect
to physical memory, the caches, the devices, and the APARA engine.
Some of the settings below are structural: they change what the
model does, e.g. how many TLB entries exist. Others are timing:
they change how long something takes, with no effect on the functional
result. Some can only be changed at compile time (edit a header,
rebuild). Others are ordinary runtime fields, given an initial value in
a configuration's own init block, but otherwise no different from any
other piece of simulation state, settable at any later point too, e.g.
with a debugger (see Examining Core State at Runtime Using
GDB) or from scripted Sitar behavior
code.
Where to set¶
Most of the timing knobs below are consolidated in one place per
configuration, alongside the wiring that connects the pieces they
belong to: Core.sitar's own init block for core_only, or
System.sitar's for core_mmu and any later configuration assembled
the same way (and Top.sitar's for apara, which has no System). Structural knobs each live in their own compile-time
header instead, listed individually below.
The timing parameters below apply only to the Sitar-timed models
(sitar_model/). The C++ only models (cpp_model/) have no notion
of cycles at all, so timing knobs have no effect on them. Structural
parameters apply to both.
Core¶
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
| Opcode latency ( DEFAULT_PER_OPCODE_DELAY, OPCODE_LATENCY_OVERRIDES) |
OpcodeLatencies.h (shared by every configuration) |
Timing | Compile time | Cycles charged per instruction, on top of any separate memory-access latency. Every opcode not listed in OPCODE_LATENCY_OVERRIDES gets the default.Range: non-negative int. |
| Memory-interface delay ( ifetchProcedure.delay, memReadProcedure.delay, memWriteProcedure.delay, atomicLoadStoreProcedure.delay, flushProcedure.delay) |
core_only: Core.sitarcore_mmu and later: that configuration's own System.sitar, e.g. System.sitar |
Timing | Runtime | One independent delay per access type (ifetch, load, store, atomic, flush), charged on the requester side, after the response is available. Models core-side interconnect/cache latency between SparcThread and whatever it talks to next (VirtualMainMemory in core_only, the MMU in core_mmu).Range: non-negative int, per field. |
NUM_THREADS_PER_CORE |
MultiThreadingConfig.h |
Structural | Compile time | Hardware threads (SMT) per core: how many independent threads, each with its own registers, trap state, and MMU registers, every core models. Only 1 is supported so far: every configuration models one thread per core. The multithreaded configurations will allow 1, 2, or 4. Range: 1. |
MMU (core_mmu and later)¶
These settings apply only to configurations that include an MMU:
core_mmu, core_mmu_devices, core_l1cache_mmu_devices, and
sparc_apara. In each of them, the runtime knobs below live in that
configuration's own System.sitar (the links below are to
core_mmu's).
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
MMU_TLB_PRESENT |
MmuConfig.h |
Structural | Compile time | Whether this build includes TLB hardware at all. With a TLB present, translations are cached. Without one, every access re-walks the page tables. The sizing constants below only matter when this is true. |
| TLB sizing ( MMU_TLB_LEVEL0_ENTRIES, MMU_TLB_LEVEL1_ENTRIES, MMU_TLB_LEVEL2_ENTRIES, MMU_TLB_LEVEL3_WAYS, MMU_TLB_LEVEL3_SETS) |
MmuConfig.h |
Structural | Compile time | Number of entries at TLB levels 0-2 (each fully associative), and the sets x ways split for level 3 (ordinary 4KB pages). Range: LEVEL0/1/2_ENTRIES, LEVEL3_WAYS: positive int. LEVEL3_SETS: power of two, 1 to 2^19. |
MMU_CONTROL_ALWAYS_CACHEABLE_BIT |
MmuConfig.h |
Structural | Compile time | Which bit of the MMU Control Register's implementation-defined SC field means "always cacheable, regardless of the PTE's own C bit" (Appendix H.3). Range: 8-23. |
MmuCore::tlbEnabled |
cpp_model: sparc_sim.cpp (MMU_TLB_ENABLED environment variable) or directlysitar_model: System.sitar (core.mmu.mmu.tlbEnabled) |
Structural | Runtime | Whether the (compiled-in) TLB is actually used for a specific run. With it false, every access re-walks the page tables regardless of what's compiled in. Range: true/false (false if MMU_TLB_PRESENT is false). |
MmuCore::atomicRMUpdates |
The multicore configuration's sparc_sim.cpp, for both models |
Structural | Runtime | Whether an R or M update re-reads the page table entry and sets its bits there, so that it cannot undo another core's change. Set by the multicore models when more than one core runs. It costs one more memory read per update in the sitar_model. Range: true/false (default false). |
core.mmu.tlbLookupDelay |
System.sitar |
Timing | Runtime | TLB tag-compare cost, charged whether the lookup hits or misses. Range: non-negative int. |
core.mmu.pageTableWalkStepDelay |
System.sitar |
Timing | Runtime | Page-table walker control overhead, charged once per level visited on a walk. Range: non-negative int. |
core.mmu.tlbFillDelay |
System.sitar |
Timing | Runtime | Cost of inserting a resolved entry into the TLB, on a fresh miss only. Range: non-negative int. |
core.mmu.registerAccessDelay |
System.sitar |
Timing | Runtime | An MMU control/context/FSR/FAR register read or write (ASI 4). Range: non-negative int. |
core.mmu.flushDelay |
System.sitar |
Timing | Runtime | A TLB invalidate (ASI 3 write). No memory access, since the TLB is write-through. Range: non-negative int. |
| MMU-to-memory interconnect delay ( core.mmu.phyMemReadProcedure.delay, core.mmu.phyMemWriteProcedure.delay) |
System.sitar |
Timing | Runtime | The interface/interconnect delay between the MMU and main memory. The actual delay is 1 plus the value set here: Sitar always adds a compulsory 1 cycle for communication over a net, on top of whatever additional delay this field specifies. Range: non-negative int, per field. |
Memory¶
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
mainMemory.delay |
core_only: Core.sitarcore_mmu and later: that configuration's own System.sitar, e.g. System.sitar |
Timing | Runtime | The memory's own service time, shared by every requester. VirtualMainMemory.delay in core_only, PhysicalMainMemory.delay in core_mmu and later.Range: non-negative int. |
Caches (core_l1cache_mmu_devices and sparc_apara)¶
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
| Cache sizes ( --icache-lines, --icache-ways, --dcache-lines, --dcache-ways) |
command line, both models (defaults in CacheConfig.h) |
Structural | Runtime | Each cache's number of 64-byte lines, and its associativity (1 = direct mapped). Range: powers of two, ways no more than lines. Default 512 lines, 1 way. |
--always-cacheable |
command line, both models (or SPARC_ALWAYS_CACHEABLE=1 in the environment) |
Structural | Runtime | Sets the MMU Control Register's "always cacheable" bit before the run, so ordinary accesses are cached with the MMU disabled. |
core.caches.icacheHitDelay, core.caches.dcacheHitDelay |
System.sitar |
Timing | Runtime | Cycles charged on an instruction or data cache hit (a data cache hit can be a read or a write). Range: non-negative int, default 0. |
core.caches.fillDelay |
System.sitar |
Timing | Runtime | Cycles charged when a line is installed, on top of the line read itself. Range: non-negative int, default 0. |
Bus and devices (core_mmu_devices and up)¶
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
bridge.delay |
System.sitar |
Timing | Runtime | Cycles the bridge adds to every physical transaction (memory and device alike), on top of the compulsory 1 cycle for each net crossed. Range: non-negative int. |
interconnect.delay |
sparc_apara's System.sitar |
Timing | Runtime | The same as bridge.delay, for sparc_apara's two-master interconnect, which replaces the bridge there. Charged to every transaction, from the core or from the APARA engine.Range: non-negative int. |
devices.registerAccessDelay |
System.sitar |
Timing | Runtime | Cycles added to every device register access. Range: non-negative int. |
timerTickDivider |
sitar_model: System.sitar (devices.devices.timerTickDivider)both models: --timer-tick-divider <n> on the command line |
Structural | Runtime | Core cycles per timer tick (the timer's clock_in). In the cpp_model a "cycle" is one instruction.Range: positive int, default 1. |
txTicksPerByte |
sitar_model: System.sitar (devices.devices.serial.txTicksPerByte) |
Timing | Runtime | Cycles a transmitted byte keeps the Serial device's Tx_full set. 0 sends a byte within the store itself.Range: non-negative int, default 1. |
| Serial I/O ( --serial-in, --serial-out, --serial-pty) |
command line, both models | Structural | Runtime | Where received bytes come from and transmitted bytes go. See the Serial Device page. |
Multicore (multicore)¶
The multicore configuration has every setting above, for each of its
cores. In its System.sitar the per-core timing knobs are set in a
loop, the same for every core (cores[i].mmu.tlbLookupDelay, and so
on). It adds the following. See The Multicore System.
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
NUM_CORES |
cpp_model: MulticoreConfig.hsitar_model: System's parameter in System.sitar |
Structural | Compile time | Number of cores, each with its own L1 caches and MMU. Range: 1 to 4, default 4. |
| Coherence ( --coherent, --no-coherent) |
both models: command line, or SPARC_COHERENT |
Structural | Runtime | Whether a write by one core removes that line from every other core's caches. See Coherence. Default: off. |
| Snoop filter ( --snoop-filter) |
both models: command line, or SPARC_SNOOP_FILTER |
Structural | Runtime | Leaves out repeated invalidates. The results of a correct program do not change, though timing can change slightly. Default: off. |
caches.invalidateLookupDelay |
sitar_model: System.sitar |
Timing | Runtime | Cycles charged to a core's next cache access for each invalidate it has received. Default: 0. |
bridge.delay |
sitar_model: System.sitar |
Timing | Runtime | Cycles the shared bus adds to every transaction, as SystemBridge's knob of the same name does.Range: non-negative int, default 0. |
| Scheduling strategy ( --schedule <spec>) |
cpp_model: command line, or SPARC_SCHEDULE |
Structural | Runtime | How the cores' instructions are interleaved: rr:K, rr-fixed:K or random:SEED[:MAXTURN]. See Execution Model.Default: rr:1. |
| Running cores ( --run-cores <m>) |
both models: command line, or SPARC_RUN_CORES |
Structural | Runtime | Runs cores 0 to m-1 only. The others stay in reset.Range: 1 to NUM_CORES, default NUM_CORES. |
| Simulation threads ( --threads <n>) |
sitar_model built with --openmp: command line, or OMP_NUM_THREADS |
Simulation only | Runtime | Number of host threads that run the simulation. The results do not depend on it. See Parallel Simulation. Range: 1 to NUM_CORES, default 1. |
| Thread map ( --thread-map <t0,t1,...>) |
sitar_model built with --openmp: command line |
Simulation only | Runtime | The host thread of each core and of the four shared modules, instead of the default division. Default: blocks of consecutive cores, the shared modules with the last thread with the fewest cores. |
APARA engine (apara and sparc_apara)¶
The engine's own settings. Each runtime knob is set in one place per
configuration: Top.sitar in apara, System.sitar in sparc_apara.
See APARA Timing Model for what each one
models.
| Parameter | Where to modify | Type | Compile time or runtime | Description |
|---|---|---|---|---|
| Instruction latencies (the latency table) |
defaults in AparaTiming.h (LatencyTable)apara: --latency-file <file> on the command line |
Timing | apara: Runtime sparc_apara: Compile time |
Cycles a bundle spends in EXEC for each class of instruction (integer add, f32 multiply, divide, ...). A bundle takes its slowest instruction's figure. See APARA Timing Model for the table and the file format. Range: positive int, per entry. |
engine.fetch.imemDelay |
apara: Top.sitarsparc_apara: System.sitar |
Timing | Runtime | Cycles between FETCH picking a thread and that thread's line of instructions arriving from IMEM. Range: non-negative int, default 0. |
engine.exec.pipelined |
same | Timing | Runtime | Whether EXEC is internally pipelined (takes a new bundle every cycle), or holds one bundle for its whole latency. Range: true/false, default true. |
engine.exec.latency.bankConflictCycles |
same | Timing | Runtime | Cycles added for each extra access, in one bundle, to the same bank of data memory. Range: non-negative int, default 1. |
engine.regfile.tokens |
same | Timing | Runtime | How many bundles may be in EXEC at once. Range: positive int, default 2. |
engine.regfile.interlock |
same | Structural (for checking only) | Runtime | Whether REGFILE holds a bundle until the registers it uses are free. false shows which reads would be hazards without the interlock (see Hazards): the results stay right (the functional state runs each thread in order), but the timing is no longer the engine's.Range: true/false, default true. |
See Performance Modeling for how to confirm a change to any of the timing knobs above actually took effect, and for the model's own reported performance measures.