Profile cold unit module loading

This commit is contained in:
udo 2026-07-16 19:57:40 +00:00
parent 91074ccec9
commit 11025ab8b3
5 changed files with 83 additions and 6 deletions

View File

@ -137,6 +137,10 @@ own import vector, memory/table-base globals, GOT globals and function-table
slots, then runs relocations, constructors, and request-pointer binding. This
removes repeated Wasmtime import-type traversal without sharing request state or
changing core-first, unit-load-order symbol resolution.
`request_perf()` reports worker module-cache hits and misses and divides a miss
into artifact lookup, wasm read, custom-section parse, serialized-module
deserialization or wasm compilation, and immutable import classification. This
keeps cold-worker module latency distinguishable without exposing source paths.
Cold module compilation and deserialization are host work, so `load_unit()`
refreshes the epoch deadline before its first guest call. Otherwise a component
whose compilation outlasted the guest CPU budget would immediately trap in the

View File

@ -104,6 +104,8 @@ RENDER(Request& context)
print("Component resolves: ", (u64)perf["component_resolve_count"].to_u64(), " / ", (u64)perf["component_resolve_us"].to_u64(), " us\n");
print("Component phases: path ", (u64)perf["component_path_us"].to_u64(), " / artifact ", (u64)perf["component_artifact_us"].to_u64(), " / load ", (u64)perf["component_load_us"].to_u64(), " / link ", (u64)perf["component_link_us"].to_u64(), " us\n");
print("Unit loads: ", (u64)perf["unit_load_count"].to_u64(), " / module ", (u64)perf["unit_module_us"].to_u64(), " / allocate ", (u64)perf["unit_allocate_us"].to_u64(), " / imports ", (u64)perf["unit_import_us"].to_u64(), " / instantiate ", (u64)perf["unit_instantiate_us"].to_u64(), " / initialize ", (u64)perf["unit_initialize_us"].to_u64(), " us\n");
print("Unit module cache: ", (u64)perf["unit_module_cache_hit_count"].to_u64(), " worker hits / ", (u64)perf["unit_module_cache_miss_count"].to_u64(), " misses / ", (u64)perf["unit_module_serialized_cache_hit_count"].to_u64(), " serialized hits / ", (u64)perf["unit_module_compile_count"].to_u64(), " wasm compiles\n");
print("Unit module phases: lookup ", (u64)perf["unit_module_lookup_us"].to_u64(), " / read ", (u64)perf["unit_module_read_us"].to_u64(), " / parse ", (u64)perf["unit_module_parse_us"].to_u64(), " / deserialize/compile ", (u64)perf["unit_module_compile_us"].to_u64(), " / classify ", (u64)perf["unit_module_classify_us"].to_u64(), " us\n");
print("Unit symbol resolves: ", (u64)perf["unit_symbol_resolve_count"].to_u64(), " / ", (u64)perf["unit_symbol_resolve_us"].to_u64(), " us\n");
print("Output buffer size: ", context.ob->str().length(), "\n");
?></pre>

View File

@ -15,6 +15,8 @@ Component resolution is divided into `component_path_us`, `component_artifact_us
Successful first loads within the request are counted by `unit_load_count` and divided into `unit_module_us`, `unit_allocate_us`, `unit_import_us`, `unit_instantiate_us`, and `unit_initialize_us`. These cover compiled-module lookup, guest memory/table allocation, import construction, Wasmtime instantiation, and relocations/constructors/request binding. Repeated handlers from an already loaded unit are excluded.
`unit_module_cache_hit_count` and `unit_module_cache_miss_count` divide module loads by the worker's compiled-module cache. A miss is further identified by `unit_module_serialized_cache_hit_count` when Wasmtime deserializes the current `.cwasm`; `unit_module_compile_count` means it fell back to compiling the `.wasm`. `unit_module_lookup_us`, `unit_module_read_us`, `unit_module_parse_us`, `unit_module_compile_us`, and `unit_module_classify_us` divide `unit_module_us` into artifact stat/cache lookup, wasm read, custom-section parse, deserialize-or-compile, and immutable import classification. The phase sum can be below the total because allocation and cache publication overhead remain in the aggregate.
`unit_symbol_resolve_count` and `unit_symbol_resolve_us` isolate function and data symbol lookup within `unit_import_us`. The remainder of import time covers import-vector construction, Wasmtime Globals, GOT function table placement, and related bindings.
Wasm FastCGI workers retain up to `MYSQL_PERSISTENT_POOL_SIZE` credential-keyed MySQL connections (default `8`; set `0` to disable). UCE calls the client library's connection-reset operation before another request receives a cached connection, clearing transactions, temporary tables, session variables, and selected databases while avoiding a new authentication handshake. Same-request leases continue to share state until request cleanup.

View File

@ -151,7 +151,8 @@ RENDER(Request& context)
check("to_u64() / to_s64() / to_f64() / to_bool()", to_u64(cfg["u64"], 7) == 42 && to_u64(cfg["bad"], 7) == 7 && to_s64(cfg["s64"], 5) == -12 && to_s64(cfg["bad"], 5) == 5 && to_f64(cfg["f64"], 1.0) > 3.49 && to_f64(cfg["bad"], 1.25) == 1.25 && to_bool(cfg["yes"], false) && !to_bool(cfg["no"], true) && !to_bool("unknown", false), var_dump(cfg));
DValue perf = request_perf();
bool perf_stable = perf["worker_pid"].to_u64() > 0 && perf["running_us"].to_f64() > 0 && perf["dispatch_us"].type != 'S' && perf["workspace_setup_us"].type != 'S' && perf["workspace_birth_us"].type != 'S' && perf["context_apply_us"].type != 'S' && perf["hostcall_count"].to_u64() > 0 && perf["hostcall_us"].type != 'S' && perf["guest_us"].to_f64() > 0 && perf["component_resolve_count"].type != 'S' && perf["component_path_us"].type != 'S' && perf["component_artifact_us"].type != 'S' && perf["component_load_us"].type != 'S' && perf["component_link_us"].type != 'S' && perf["unit_load_count"].to_u64() > 0 && perf["unit_module_us"].type != 'S' && perf["unit_allocate_us"].type != 'S' && perf["unit_import_us"].type != 'S' && perf["unit_symbol_resolve_count"].type != 'S' && perf["unit_symbol_resolve_us"].type != 'S' && perf["unit_instantiate_us"].type != 'S' && perf["unit_initialize_us"].type != 'S';
bool unit_module_profile_valid = perf["unit_module_cache_hit_count"].to_u64() + perf["unit_module_cache_miss_count"].to_u64() == perf["unit_load_count"].to_u64() && perf["unit_module_serialized_cache_hit_count"].to_u64() + perf["unit_module_compile_count"].to_u64() == perf["unit_module_cache_miss_count"].to_u64() && perf["unit_module_lookup_us"].type != 'S' && perf["unit_module_read_us"].type != 'S' && perf["unit_module_parse_us"].type != 'S' && perf["unit_module_compile_us"].type != 'S' && perf["unit_module_classify_us"].type != 'S';
bool perf_stable = unit_module_profile_valid && perf["worker_pid"].to_u64() > 0 && perf["running_us"].to_f64() > 0 && perf["dispatch_us"].type != 'S' && perf["workspace_setup_us"].type != 'S' && perf["workspace_birth_us"].type != 'S' && perf["context_apply_us"].type != 'S' && perf["hostcall_count"].to_u64() > 0 && perf["hostcall_us"].type != 'S' && perf["guest_us"].to_f64() > 0 && perf["component_resolve_count"].type != 'S' && perf["component_path_us"].type != 'S' && perf["component_artifact_us"].type != 'S' && perf["component_load_us"].type != 'S' && perf["component_link_us"].type != 'S' && perf["unit_load_count"].to_u64() > 0 && perf["unit_module_us"].type != 'S' && perf["unit_allocate_us"].type != 'S' && perf["unit_import_us"].type != 'S' && perf["unit_symbol_resolve_count"].type != 'S' && perf["unit_symbol_resolve_us"].type != 'S' && perf["unit_instantiate_us"].type != 'S' && perf["unit_initialize_us"].type != 'S';
u64 profiled_hostcalls = perf["hostcall_count"].to_u64();
f64 profiled_hostcall_us = perf["hostcall_us"].to_f64();
u64 profiled_components = perf["component_resolve_count"].to_u64();

View File

@ -103,6 +103,17 @@ struct WasmUnitModule
size_t got_memory_imports = 0;
};
struct WasmUnitModuleLoadProfile
{
bool cache_hit = false;
bool serialized_cache_hit = false;
u64 lookup_us = 0;
u64 read_us = 0;
u64 parse_us = 0;
u64 compile_us = 0;
u64 classify_us = 0;
};
struct WasmWorkerConfig
{
String core_wasm_path = "bin/wasm/core.wasm";
@ -141,6 +152,15 @@ struct WasmRequestProfile
u64 component_link_total_us = 0;
u64 unit_load_count = 0;
u64 unit_module_total_us = 0;
u64 unit_module_cache_hit_count = 0;
u64 unit_module_cache_miss_count = 0;
u64 unit_module_serialized_cache_hit_count = 0;
u64 unit_module_compile_count = 0;
u64 unit_module_lookup_total_us = 0;
u64 unit_module_read_total_us = 0;
u64 unit_module_parse_total_us = 0;
u64 unit_module_compile_total_us = 0;
u64 unit_module_classify_total_us = 0;
u64 unit_allocate_total_us = 0;
u64 unit_import_total_us = 0;
u64 unit_symbol_resolve_count = 0;
@ -730,7 +750,8 @@ public:
String core_cached_path = cached_wasm_path(cfg.core_wasm_path);
String compile_error;
auto compiled_or_cached = load_or_compile_cached_module(engine, core_cached_path, cfg.core_wasm_path, bytes, compile_error);
bool serialized_cache_hit = false;
auto compiled_or_cached = load_or_compile_cached_module(engine, core_cached_path, cfg.core_wasm_path, bytes, compile_error, serialized_cache_hit);
if(!compiled_or_cached)
return("core module compile failed: " + compile_error);
core_module.emplace(std::move(*compiled_or_cached));
@ -747,20 +768,28 @@ public:
return(cfg.cache_root + source_path + ".wasm");
}
std::shared_ptr<WasmUnitModule> unit_module(const String& source_path, String& error)
std::shared_ptr<WasmUnitModule> unit_module(const String& source_path, String& error, WasmUnitModuleLoadProfile& profile)
{
auto lookup_start = std::chrono::steady_clock::now();
String wasm_path = unit_wasm_path(source_path);
struct stat st;
if(stat(wasm_path.c_str(), &st) != 0)
{
profile.lookup_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - lookup_start).count();
error = "no wasm artifact for " + source_path + " (expected " + wasm_path + ")";
return(nullptr);
}
auto cached = module_cache.find(wasm_path);
u64 modified_ns = (u64)st.st_mtim.tv_sec * 1000000000ull + (u64)st.st_mtim.tv_nsec;
u64 changed_ns = (u64)st.st_ctim.tv_sec * 1000000000ull + (u64)st.st_ctim.tv_nsec;
profile.lookup_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - lookup_start).count();
if(cached != module_cache.end() && cached->second->modified_ns == modified_ns && cached->second->changed_ns == changed_ns && cached->second->size == (u64)st.st_size)
{
profile.cache_hit = true;
return(cached->second);
}
auto unit = std::make_shared<WasmUnitModule>();
unit->source_path = source_path;
@ -769,16 +798,26 @@ public:
unit->changed_ns = changed_ns;
unit->size = (u64)st.st_size;
std::vector<u8> bytes;
auto read_start = std::chrono::steady_clock::now();
if(!wasm_read_file(wasm_path, bytes))
{
profile.read_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - read_start).count();
error = "cannot read " + wasm_path;
return(nullptr);
}
profile.read_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - read_start).count();
auto parse_start = std::chrono::steady_clock::now();
if(!wasm_parse_sections(bytes, unit->dylink, unit->abi, error))
{
profile.parse_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - parse_start).count();
error = wasm_path + ": " + error;
return(nullptr);
}
profile.parse_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - parse_start).count();
if(!unit->dylink.found)
{
error = wasm_path + ": missing dylink.0 mem_info (not a PIC side module)";
@ -791,13 +830,17 @@ public:
}
String unit_cached_path = cached_wasm_path(wasm_path);
String compile_error;
auto compiled_or_cached = load_or_compile_cached_module(engine, unit_cached_path, wasm_path, bytes, compile_error);
auto compile_start = std::chrono::steady_clock::now();
auto compiled_or_cached = load_or_compile_cached_module(engine, unit_cached_path, wasm_path, bytes, compile_error, profile.serialized_cache_hit);
profile.compile_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - compile_start).count();
if(!compiled_or_cached)
{
error = wasm_path + ": compile failed: " + compile_error;
return(nullptr);
}
unit->module.emplace(std::move(*compiled_or_cached));
auto classify_start = std::chrono::steady_clock::now();
for(auto import_type : unit->module->imports())
{
String mod_name(import_type.module());
@ -831,6 +874,8 @@ public:
}
unit->imports.push_back({ kind, std::move(name) });
}
profile.classify_us = (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - classify_start).count();
module_cache[wasm_path] = unit;
return(unit);
}
@ -844,8 +889,9 @@ private:
}
static std::optional<wasmtime::Module> load_or_compile_cached_module(wasmtime::Engine& engine, const String& cached_path, const String& wasm_path,
std::vector<u8>& bytes, String& compile_error)
std::vector<u8>& bytes, String& compile_error, bool& serialized_cache_hit)
{
serialized_cache_hit = false;
struct stat wasm_stat;
struct stat cached_stat;
if(stat(cached_path.c_str(), &cached_stat) == 0 && stat(wasm_path.c_str(), &wasm_stat) == 0)
@ -855,7 +901,10 @@ private:
{
auto deserialized = wasmtime::Module::deserialize_file(engine, cached_path);
if(deserialized)
{
serialized_cache_hit = true;
return(deserialized.ok());
}
}
}
@ -1338,9 +1387,19 @@ private:
unit_load_count++;
String error;
auto module_start = std::chrono::steady_clock::now();
auto mod = worker.unit_module(source_path, error);
WasmUnitModuleLoadProfile module_profile;
auto mod = worker.unit_module(source_path, error, module_profile);
unit_module_total_us += (u64)std::chrono::duration_cast<std::chrono::microseconds>(
std::chrono::steady_clock::now() - module_start).count();
unit_module_cache_hit_count += module_profile.cache_hit ? 1 : 0;
unit_module_cache_miss_count += module_profile.cache_hit ? 0 : 1;
unit_module_serialized_cache_hit_count += module_profile.serialized_cache_hit ? 1 : 0;
unit_module_compile_count += !module_profile.cache_hit && !module_profile.serialized_cache_hit ? 1 : 0;
unit_module_lookup_total_us += module_profile.lookup_us;
unit_module_read_total_us += module_profile.read_us;
unit_module_parse_total_us += module_profile.parse_us;
unit_module_compile_total_us += module_profile.compile_us;
unit_module_classify_total_us += module_profile.classify_us;
if(!mod)
return(error);
// Compiling/deserializing a cold module is host work. Refresh the guest
@ -1977,6 +2036,15 @@ private:
response["component_link_us"] = (f64)self->component_link_total_us;
response["unit_load_count"] = (f64)self->unit_load_count;
response["unit_module_us"] = (f64)self->unit_module_total_us;
response["unit_module_cache_hit_count"] = (f64)self->unit_module_cache_hit_count;
response["unit_module_cache_miss_count"] = (f64)self->unit_module_cache_miss_count;
response["unit_module_serialized_cache_hit_count"] = (f64)self->unit_module_serialized_cache_hit_count;
response["unit_module_compile_count"] = (f64)self->unit_module_compile_count;
response["unit_module_lookup_us"] = (f64)self->unit_module_lookup_total_us;
response["unit_module_read_us"] = (f64)self->unit_module_read_total_us;
response["unit_module_parse_us"] = (f64)self->unit_module_parse_total_us;
response["unit_module_compile_us"] = (f64)self->unit_module_compile_total_us;
response["unit_module_classify_us"] = (f64)self->unit_module_classify_total_us;
response["unit_allocate_us"] = (f64)self->unit_allocate_total_us;
response["unit_import_us"] = (f64)self->unit_import_total_us;
response["unit_symbol_resolve_count"] = (f64)self->unit_symbol_resolve_count;