perf(ecstore): reuse erasure codecs on GET paths (#6074)

* perf(ecstore): share legacy SIMD workspaces

Reuse legacy Reed-Solomon encoder and decoder workspaces across Erasure instances with the same shard layout while keeping active codecs request-exclusive.

Co-Authored-By: heihutu <heihutu@gmail.com>

* perf(ecstore): reuse GET erasure shells and scratch buffers

Co-Authored-By: heihutu <heihutu@gmail.com>

* test(ecstore): satisfy concurrent codec lint

Co-Authored-By: heihutu <heihutu@gmail.com>

* perf(ecstore): bound cached legacy workspaces

Co-Authored-By: heihutu <heihutu@gmail.com>

* perf(ecstore): cap retained legacy codec memory

Co-Authored-By: heihutu <heihutu@gmail.com>

---------

Co-authored-by: heihutu <heihutu@gmail.com>
This commit is contained in:
houseme
2026-08-14 01:51:42 +08:00
committed by GitHub
parent e16c07b9cd
commit 6178083985
5 changed files with 472 additions and 72 deletions
+194 -40
View File
@@ -71,10 +71,16 @@ impl EncodedBlock {
const MODERN_MAX_TOTAL_SHARDS: usize = <reed_solomon_erasure::galois_8::Field as reed_solomon_erasure::Field>::ORDER;
const MODERN_REED_SOLOMON_CACHE_MAX_ENTRIES: usize = 64;
const LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES: usize = 16;
// Vec growth may retain twice the requested logical length. Keeping the logical
// workspace at half the budget bounds each cached workspace's shard allocation to 1 MiB.
const LEGACY_REED_SOLOMON_CACHE_MAX_LOGICAL_SHARD_BYTES_PER_WORKSPACE: usize = 512 * 1024;
type ModernReedSolomonCache = RwLock<HashMap<(usize, usize), Arc<ReedSolomon>>>;
type LegacyReedSolomonCache = RwLock<HashMap<(usize, usize), Arc<LegacyReedSolomonEncoder>>>;
static MODERN_REED_SOLOMON_CACHE: OnceLock<ModernReedSolomonCache> = OnceLock::new();
static LEGACY_REED_SOLOMON_CACHE: OnceLock<LegacyReedSolomonCache> = OnceLock::new();
/// Errors returned when constructing an [`Erasure`] codec.
#[derive(Debug, thiserror::Error)]
@@ -141,43 +147,61 @@ pub fn calc_shard_size_legacy(block_size: usize, data_shards: usize) -> usize {
struct LegacyReedSolomonEncoder {
data_shards: usize,
parity_shards: usize,
encoder_cache: std::sync::RwLock<Option<reed_solomon_simd::ReedSolomonEncoder>>,
decoder_cache: std::sync::RwLock<Option<reed_solomon_simd::ReedSolomonDecoder>>,
}
impl Clone for LegacyReedSolomonEncoder {
fn clone(&self) -> Self {
Self {
data_shards: self.data_shards,
parity_shards: self.parity_shards,
encoder_cache: std::sync::RwLock::new(None),
decoder_cache: std::sync::RwLock::new(None),
}
}
cache_workspaces: bool,
encoder_cache: RwLock<Option<reed_solomon_simd::ReedSolomonEncoder>>,
decoder_cache: RwLock<Option<reed_solomon_simd::ReedSolomonDecoder>>,
}
impl LegacyReedSolomonEncoder {
fn new(_data_shards: usize, _parity_shards: usize) -> io::Result<Self> {
fn new(data_shards: usize, parity_shards: usize) -> io::Result<Self> {
Self::with_workspace_cache(data_shards, parity_shards, false)
}
fn with_workspace_cache(data_shards: usize, parity_shards: usize, cache_workspaces: bool) -> io::Result<Self> {
Ok(Self {
data_shards: _data_shards,
parity_shards: _parity_shards,
encoder_cache: std::sync::RwLock::new(None),
decoder_cache: std::sync::RwLock::new(None),
data_shards,
parity_shards,
cache_workspaces,
encoder_cache: RwLock::new(None),
decoder_cache: RwLock::new(None),
})
}
fn logical_shard_bytes_upper_bound(&self, shard_len: usize) -> Option<usize> {
let aligned_shard_len = shard_len.checked_add(63)?.checked_div(64)?.checked_mul(64)?;
let high_rate_decoder_work_count = self
.parity_shards
.checked_next_power_of_two()?
.checked_add(self.data_shards)?
.checked_next_power_of_two()?;
let low_rate_decoder_work_count = self
.data_shards
.checked_next_power_of_two()?
.checked_add(self.parity_shards)?
.checked_next_power_of_two()?;
aligned_shard_len.checked_mul(high_rate_decoder_work_count.max(low_rate_decoder_work_count))
}
fn should_cache_workspace(&self, shard_len: usize) -> bool {
self.cache_workspaces
&& self
.logical_shard_bytes_upper_bound(shard_len)
.is_some_and(|bytes| bytes <= LEGACY_REED_SOLOMON_CACHE_MAX_LOGICAL_SHARD_BYTES_PER_WORKSPACE)
}
fn encode(&self, shards: SmallVec<[&mut [u8]; 16]>) -> io::Result<()> {
let mut shards_vec: Vec<&mut [u8]> = shards.into_vec();
if shards_vec.is_empty() {
return Ok(());
}
let shard_len = shards_vec[0].len();
let cached_encoder = self
.encoder_cache
.write()
.map_err(|_| io::Error::other("Failed to acquire encoder cache lock"))?
.take();
let mut encoder = {
let mut cache_guard = self
.encoder_cache
.write()
.map_err(|_| io::Error::other("Failed to acquire encoder cache lock"))?;
match cache_guard.take() {
match cached_encoder {
Some(mut cached) => {
if cached.reset(self.data_shards, self.parity_shards, shard_len).is_err() {
reed_solomon_simd::ReedSolomonEncoder::new(self.data_shards, self.parity_shards, shard_len)
@@ -204,10 +228,15 @@ impl LegacyReedSolomonEncoder {
}
}
drop(result);
*self
.encoder_cache
.write()
.map_err(|_| io::Error::other("Failed to return encoder to cache"))? = Some(encoder);
if self.should_cache_workspace(shard_len) {
let mut cache = self
.encoder_cache
.write()
.map_err(|_| io::Error::other("Failed to return encoder to cache"))?;
if cache.is_none() {
*cache = Some(encoder);
}
}
Ok(())
}
@@ -221,13 +250,13 @@ impl LegacyReedSolomonEncoder {
.find_map(|s| s.as_ref().map(|v| v.len()))
.ok_or_else(|| io::Error::other("No valid shards found for reconstruction"))?;
let cached_decoder = self
.decoder_cache
.write()
.map_err(|_| io::Error::other("Failed to acquire decoder cache lock"))?
.take();
let mut decoder = {
let mut cache_guard = self
.decoder_cache
.write()
.map_err(|_| io::Error::other("Failed to acquire decoder cache lock"))?;
match cache_guard.take() {
match cached_decoder {
Some(mut cached_decoder) => {
if let Err(e) = cached_decoder.reset(self.data_shards, self.parity_shards, shard_len) {
warn!("Failed to reset SIMD decoder: {:?}, creating new one", e);
@@ -274,10 +303,15 @@ impl LegacyReedSolomonEncoder {
drop(result);
*self
.decoder_cache
.write()
.map_err(|_| io::Error::other("Failed to return decoder to cache"))? = Some(decoder);
if self.should_cache_workspace(shard_len) {
let mut cache = self
.decoder_cache
.write()
.map_err(|_| io::Error::other("Failed to return decoder to cache"))?;
if cache.is_none() {
*cache = Some(decoder);
}
}
Ok(())
}
@@ -435,6 +469,39 @@ fn cached_modern_reed_solomon(data_shards: usize, parity_shards: usize) -> Resul
Ok(encoder)
}
fn cached_legacy_reed_solomon(data_shards: usize, parity_shards: usize) -> io::Result<Arc<LegacyReedSolomonEncoder>> {
let cache = LEGACY_REED_SOLOMON_CACHE.get_or_init(|| RwLock::new(HashMap::new()));
cached_legacy_reed_solomon_in(cache, data_shards, parity_shards)
}
fn cached_legacy_reed_solomon_in(
cache: &LegacyReedSolomonCache,
data_shards: usize,
parity_shards: usize,
) -> io::Result<Arc<LegacyReedSolomonEncoder>> {
let key = (data_shards, parity_shards);
if let Some(encoder) = cache
.read()
.unwrap_or_else(|poisoned| poisoned.into_inner())
.get(&key)
.cloned()
{
return Ok(encoder);
}
let mut cache = cache.write().unwrap_or_else(|poisoned| poisoned.into_inner());
if let Some(existing) = cache.get(&key) {
return Ok(Arc::clone(existing));
}
if cache.len() < LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES {
let encoder = Arc::new(LegacyReedSolomonEncoder::with_workspace_cache(data_shards, parity_shards, true)?);
cache.insert(key, Arc::clone(&encoder));
return Ok(encoder);
}
drop(cache);
Ok(Arc::new(LegacyReedSolomonEncoder::new(data_shards, parity_shards)?))
}
fn encode_parity_shards<F>(shards: &mut [Option<Vec<u8>>], data_shards: usize, parity_shards: usize, encode: F) -> io::Result<()>
where
F: FnOnce(SmallVec<[&mut [u8]; 16]>) -> io::Result<()>,
@@ -551,7 +618,7 @@ pub struct Erasure {
pub data_shards: usize,
pub parity_shards: usize,
encoder: Option<ReedSolomonEncoder>,
legacy_encoder: Option<LegacyReedSolomonEncoder>,
legacy_encoder: Option<Arc<LegacyReedSolomonEncoder>>,
pub block_size: usize,
uses_legacy: bool,
_id: Uuid,
@@ -687,7 +754,7 @@ impl Erasure {
let legacy_encoder = if uses_legacy && parity_shards > 0 {
Some(
LegacyReedSolomonEncoder::new(data_shards, parity_shards)
cached_legacy_reed_solomon(data_shards, parity_shards)
.map_err(|source| ErasureConstructionError::LegacyEncoder { source })?,
)
} else {
@@ -1405,7 +1472,7 @@ mod tests {
assert_eq!(cloned.block_size, legacy.block_size);
assert!(cloned.uses_legacy);
let data = b"legacy clone should keep independent SIMD caches";
let data = b"legacy clone should preserve SIMD codec behavior";
let encoded = cloned.encode_data(data).expect("legacy clone should encode");
let mut shards = optional_shards(&encoded);
shards[0] = None;
@@ -1413,6 +1480,93 @@ mod tests {
assert_eq!(recover_data(&shards, cloned.data_shards, data.len()), data);
}
#[test]
fn legacy_codecs_share_process_cache_across_erasure_instances() {
let first = Erasure::new_with_options(6, 3, 64, true)
.legacy_encoder
.expect("legacy codec should be initialized");
let second = Erasure::new_with_options(6, 3, 128, true)
.legacy_encoder
.expect("same legacy shard layout should be initialized");
assert!(Arc::ptr_eq(&first, &second));
}
#[test]
fn legacy_workspace_cache_rejects_oversize_buffers_and_isolates_layouts() {
let four_plus_two = Erasure::new_with_options(4, 2, 64, true)
.legacy_encoder
.expect("legacy codec should be initialized");
let four_plus_one = Erasure::new_with_options(4, 1, 64, true)
.legacy_encoder
.expect("distinct parity layout should be initialized");
let three_plus_two = Erasure::new_with_options(3, 2, 64, true)
.legacy_encoder
.expect("distinct data layout should be initialized");
assert!(!Arc::ptr_eq(&four_plus_two, &four_plus_one));
assert!(!Arc::ptr_eq(&four_plus_two, &three_plus_two));
assert_eq!(four_plus_two.logical_shard_bytes_upper_bound(64 * 1024), Some(512 * 1024));
assert!(four_plus_two.should_cache_workspace(64 * 1024));
assert!(!four_plus_two.should_cache_workspace(64 * 1024 + 1));
let nine_plus_seven =
LegacyReedSolomonEncoder::with_workspace_cache(9, 7, true).expect("9+7 legacy codec should construct");
assert_eq!(nine_plus_seven.logical_shard_bytes_upper_bound(16 * 1024), Some(512 * 1024));
assert!(nine_plus_seven.should_cache_workspace(16 * 1024));
assert!(!nine_plus_seven.should_cache_workspace(16 * 1024 + 1));
let uncached = LegacyReedSolomonEncoder::new(4, 2).expect("uncached legacy codec should construct");
assert!(!uncached.should_cache_workspace(64));
}
#[test]
fn saturated_legacy_codec_cache_does_not_retain_more_workspaces() {
let cache = RwLock::new(HashMap::new());
for parity_shards in 1..=LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES {
let cached =
cached_legacy_reed_solomon_in(&cache, 32, parity_shards).expect("cacheable legacy codec should construct");
assert!(cached.cache_workspaces);
}
let uncached =
cached_legacy_reed_solomon_in(&cache, 31, 1).expect("uncached legacy codec should construct after saturation");
assert!(!uncached.cache_workspaces);
assert_eq!(
cache.read().expect("cache lock should remain healthy").len(),
LEGACY_REED_SOLOMON_CACHE_MAX_ENTRIES
);
}
#[test]
fn concurrent_legacy_codecs_preserve_byte_exact_results() {
let barrier = Arc::new(std::sync::Barrier::new(2));
let payloads = [vec![0x35; 257], vec![0xca; 1025]];
std::thread::scope(|scope| {
let handles = payloads.each_ref().map(|payload| {
let barrier = Arc::clone(&barrier);
scope.spawn(move || {
let erasure = Erasure::new_with_options(6, 3, 2048, true);
barrier.wait();
let encoded = erasure.encode_data(payload).expect("concurrent legacy encode should succeed");
barrier.wait();
let mut shards = optional_shards(&encoded);
shards[0] = None;
erasure
.decode_data(&mut shards)
.expect("concurrent legacy decode should reconstruct the missing shard");
recover_data(&shards, erasure.data_shards, payload.len())
})
});
for (handle, payload) in handles.into_iter().zip(payloads.iter()) {
assert_eq!(handle.join().expect("concurrent legacy codec worker should not panic"), *payload);
}
});
}
#[test]
fn legacy_verify_reports_invalid_empty_valid_and_corrupt_parity_sets() {
let legacy = LegacyReedSolomonEncoder::new(2, 2).expect("legacy encoder should construct");