GC auto increase memory limit on inefficient GC (vercel/turborepo#8687)

### Description

When GC is inefficient and can't collect enough memory, increase the
provided memory limit automatically.
We rather want to consume more memory than requested then crashing or
hanging the compilation due to GC.

crates/turbo-tasks-memory/src/gc.rs

@@ -72,18 +72,24 @@ const MAX_TASKS_PER_OLD_GENERATION: usize = 200_000;
 const PERCENTAGE_TO_COLLECT: usize = 30;
 const TASK_BASE_MEMORY_USAGE: usize = 1_000;
 const TASK_BASE_COMPUTE_DURATION_IN_MICROS: u64 = 1_000;
-pub const PERCENTAGE_TARGET_MEMORY: usize = 88;
-pub const PERCENTAGE_IDLE_TARGET_MEMORY: usize = 75;
+pub const PERCENTAGE_MIN_TARGET_MEMORY: usize = 70;
+pub const PERCENTAGE_MAX_TARGET_MEMORY: usize = 75;
+pub const PERCENTAGE_MIN_IDLE_TARGET_MEMORY: usize = 55;
+pub const PERCENTAGE_MAX_IDLE_TARGET_MEMORY: usize = 60;
+pub const MAX_GC_STEPS: usize = 100;
 
 struct OldGeneration {
  tasks: Vec<TaskId>,
  generation: NonZeroU32,
 }
 
+#[derive(Default)]
 struct ProcessGenerationResult {
+ old_generations: usize,
  priority: Option<GcPriority>,
  content_dropped_count: usize,
  unloaded_count: usize,
+ already_unloaded_count: usize,
 }
 
 struct ProcessDeactivationsResult {
@@ -222,22 +228,22 @@ impl GcQueue {
  &self,
  backend: &MemoryBackend,
  turbo_tasks: &dyn TurboTasksBackendApi<MemoryBackend>,
- ) -> ProcessGenerationResult {
- let old_generation = {
+ ) -> Option<ProcessGenerationResult> {
+ let old_generation_info = {
  let guard = &mut self.generations.lock();
- guard.pop_back()
+ let len = guard.len();
+ guard.pop_back().map(|g| (g, len))
  };
- let Some(OldGeneration {
- mut tasks,
- generation,
- }) = old_generation
+ let Some((
+ OldGeneration {
+ mut tasks,
+ generation,
+ },
+ old_generations,
+ )) = old_generation_info
  else {
  // No old generation to process
- return ProcessGenerationResult {
- priority: None,
- content_dropped_count: 0,
- unloaded_count: 0,
- };
+ return None;
  };
  // Check all tasks for the correct generation
  let mut indices = Vec::with_capacity(tasks.len());
@@ -256,11 +262,10 @@ impl GcQueue {
 
  if indices.is_empty() {
  // No valid tasks in old generation to process
- return ProcessGenerationResult {
- priority: None,
- content_dropped_count: 0,
- unloaded_count: 0,
- };
+ return Some(ProcessGenerationResult {
+ old_generations,
+ ..ProcessGenerationResult::default()
+ });
  }
 
  // Sorting based on sort_by_cached_key from std lib
@@ -316,6 +321,7 @@ impl GcQueue {
  // GC the tasks
  let mut content_dropped_count = 0;
  let mut unloaded_count = 0;
+ let mut already_unloaded_count = 0;
  for task in tasks[..tasks_to_collect].iter() {
  backend.with_task(*task, |task| {
  match task.run_gc(generation, self, backend, turbo_tasks) {
@@ -327,26 +333,31 @@ impl GcQueue {
  GcResult::Unloaded => {
  unloaded_count += 1;
  }
+ GcResult::AlreadyUnloaded => {
+ already_unloaded_count += 1;
+ }
  }
  });
  }
 
- ProcessGenerationResult {
+ Some(ProcessGenerationResult {
+ old_generations,
  priority: Some(max_priority),
  content_dropped_count,
  unloaded_count,
- }
+ already_unloaded_count,
+ })
  }
 
- /// Run garbage collection on the queue.
+ /// Run garbage collection on the queue. Returns true, if some progress has
+ /// been made. Returns the number of old generations.
  pub fn run_gc(
  &self,
  backend: &MemoryBackend,
  turbo_tasks: &dyn TurboTasksBackendApi<MemoryBackend>,
- ) -> Option<(GcPriority, usize)> {
+ ) -> Option<usize> {
  let span = tracing::trace_span!(
- parent: None,
- "garbage collection",
+ "garbage collection step",
  priority = Empty,
  deactivations_count = Empty,
  content_dropped_count = Empty,
@@ -359,21 +370,27 @@ impl GcQueue {
  count: deactivations_count,
  } = self.process_deactivations(backend, turbo_tasks);
 
- let ProcessGenerationResult {
+ if let Some(ProcessGenerationResult {
+ old_generations,
  priority,
  content_dropped_count,
  unloaded_count,
- } = self.process_old_generation(backend, turbo_tasks);
+ already_unloaded_count,
+ }) = self.process_old_generation(backend, turbo_tasks)
+ {
+ span.record("deactivations_count", deactivations_count);
+ span.record("content_dropped_count", content_dropped_count);
+ span.record("unloaded_count", unloaded_count);
+ span.record("already_unloaded_count", already_unloaded_count);
+ if let Some(priority) = &priority {
+ span.record("priority", debug(priority));
+ } else {
+ span.record("priority", "");
+ }
 
- span.record("deactivations_count", deactivations_count);
- span.record("content_dropped_count", content_dropped_count);
- span.record("unloaded_count", unloaded_count);
- if let Some(priority) = &priority {
- span.record("priority", debug(priority));
+ Some(old_generations)
  } else {
- span.record("priority", "");
+ (deactivations_count > 0).then_some(0)
  }
-
- priority.map(|p| (p, content_dropped_count))
  }
 }

crates/turbo-tasks-memory/src/memory_backend.rs

@@ -6,7 +6,7 @@ use std::{
  num::NonZeroU32,
  pin::Pin,
  sync::{
- atomic::{AtomicBool, Ordering},
+ atomic::{AtomicBool, AtomicUsize, Ordering},
  Arc,
  },
  time::Duration,
@@ -31,7 +31,10 @@ use turbo_tasks::{
 
 use crate::{
  edges_set::{TaskEdge, TaskEdgesSet},
- gc::{GcQueue, PERCENTAGE_IDLE_TARGET_MEMORY, PERCENTAGE_TARGET_MEMORY},
+ gc::{
+ GcQueue, MAX_GC_STEPS, PERCENTAGE_MAX_IDLE_TARGET_MEMORY, PERCENTAGE_MAX_TARGET_MEMORY,
+ PERCENTAGE_MIN_IDLE_TARGET_MEMORY, PERCENTAGE_MIN_TARGET_MEMORY,
+ },
  output::Output,
  task::{ReadCellError, Task, DEPENDENCIES_TO_TRACK},
  task_statistics::TaskStatisticsApi,
@@ -47,7 +50,7 @@ pub struct MemoryBackend {
  backend_job_id_factory: IdFactoryWithReuse<BackendJobId>,
  task_cache:
  DashMap<Arc<PreHashed<PersistentTaskType>>, TaskId, BuildHasherDefault<PassThroughHash>>,
- memory_limit: usize,
+ memory_limit: AtomicUsize,
  gc_queue: Option<GcQueue>,
  idle_gc_active: AtomicBool,
  task_statistics: TaskStatisticsApi,
@@ -70,7 +73,7 @@ impl MemoryBackend {
  (std::thread::available_parallelism().map_or(1, usize::from) * 32)
  .next_power_of_two(),
  ),
- memory_limit,
+ memory_limit: AtomicUsize::new(memory_limit),
  gc_queue: (memory_limit != usize::MAX).then(GcQueue::new),
  idle_gc_active: AtomicBool::new(false),
  task_statistics: TaskStatisticsApi::default(),
@@ -141,27 +144,78 @@ impl MemoryBackend {
  ) -> bool {
  if let Some(gc_queue) = &self.gc_queue {
  let mut did_something = false;
- loop {
- let mem_limit = self.memory_limit;
-
- let usage = turbo_tasks_malloc::TurboMalloc::memory_usage();
- let target = if idle {
- mem_limit * PERCENTAGE_IDLE_TARGET_MEMORY / 100
+ let mut remaining_generations = 0;
+ let mut mem_limit = self.memory_limit.load(Ordering::Relaxed);
+ let mut span = None;
+ 'outer: loop {
+ let mut collected_generations = 0;
+ let (min, max) = if idle {
+ (
+ mem_limit * PERCENTAGE_MIN_IDLE_TARGET_MEMORY / 100,
+ mem_limit * PERCENTAGE_MAX_IDLE_TARGET_MEMORY / 100,
+ )
  } else {
- mem_limit * PERCENTAGE_TARGET_MEMORY / 100
+ (
+ mem_limit * PERCENTAGE_MIN_TARGET_MEMORY / 100,
+ mem_limit * PERCENTAGE_MAX_TARGET_MEMORY / 100,
+ )
  };
- if usage < target {
- return did_something;
- }
-
- let collected = gc_queue.run_gc(self, turbo_tasks);
-
- // Collecting less than 100 tasks is not worth it
- if !collected.map_or(false, |(_, count)| count > 100) {
- return true;
+ let mut target = max;
+ let mut counter = 0;
+ loop {
+ let usage = turbo_tasks_malloc::TurboMalloc::memory_usage();
+ if usage < target {
+ return did_something;
+ }
+ target = min;
+ if span.is_none() {
+ span =
+ Some(tracing::trace_span!(parent: None, "garbage collection", usage));
+ }
+
+ let progress = gc_queue.run_gc(self, turbo_tasks);
+
+ if progress.is_some() {
+ did_something = true;
+ }
+
+ if let Some(g) = progress {
+ remaining_generations = g;
+ if g > 0 {
+ collected_generations += 1;
+ }
+ }
+
+ counter += 1;
+ if counter > MAX_GC_STEPS
+ || collected_generations > remaining_generations
+ || progress.is_none()
+ {
+ let new_mem_limit = mem_limit * 4 / 3;
+ if self
+ .memory_limit
+ .compare_exchange(
+ mem_limit,
+ new_mem_limit,
+ Ordering::Relaxed,
+ Ordering::Relaxed,
+ )
+ .is_ok()
+ {
+ println!(
+ "Ineffective GC, increasing memory limit {} MB -> {} MB",
+ mem_limit / 1024 / 1024,
+ new_mem_limit / 1024 / 1024
+ );
+ mem_limit = new_mem_limit;
+ } else {
+ mem_limit = self.memory_limit.load(Ordering::Relaxed);
+ }
+ continue 'outer;
+ }
+
+ did_something = true;
  }
-
- did_something = true;
  }
  }
  false

crates/turbo-tasks-memory/src/task.rs

@@ -455,6 +455,7 @@ pub enum GcResult {
  /// from the graph and only makes sense when the task isn't currently
  /// active.
  Unloaded,
+ AlreadyUnloaded,
 }
 
 pub enum ReadCellError {
@@ -1818,9 +1819,9 @@ impl Task {
  }
  TaskMetaStateWriteGuard::Partial(mut state) => {
  state.aggregation_node.shrink_to_fit();
- GcResult::Unloaded
+ GcResult::AlreadyUnloaded
  }
- TaskMetaStateWriteGuard::Unloaded(_) => GcResult::Unloaded,
+ TaskMetaStateWriteGuard::Unloaded(_) => GcResult::AlreadyUnloaded,
  TaskMetaStateWriteGuard::TemporaryFiller => unreachable!(),
  }
  }