rocksdb/cache/cache_key.cc

272 lines
13 KiB
C++
Raw Normal View History

New stable, fixed-length cache keys (#9126) Summary: This change standardizes on a new 16-byte cache key format for block cache (incl compressed and secondary) and persistent cache (but not table cache and row cache). The goal is a really fast cache key with practically ideal stability and uniqueness properties without external dependencies (e.g. from FileSystem). A fixed key size of 16 bytes should enable future optimizations to the concurrent hash table for block cache, which is a heavy CPU user / bottleneck, but there appears to be measurable performance improvement even with no changes to LRUCache. This change replaces a lot of disjointed and ugly code handling cache keys with calls to a simple, clean new internal API (cache_key.h). (Preserving the old cache key logic under an option would be very ugly and likely negate the performance gain of the new approach. Complete replacement carries some inherent risk, but I think that's acceptable with sufficient analysis and testing.) The scheme for encoding new cache keys is complicated but explained in cache_key.cc. Also: EndianSwapValue is moved to math.h to be next to other bit operations. (Explains some new include "math.h".) ReverseBits operation added and unit tests added to hash_test for both. Fixes https://github.com/facebook/rocksdb/issues/7405 (presuming a root cause) Pull Request resolved: https://github.com/facebook/rocksdb/pull/9126 Test Plan: ### Basic correctness Several tests needed updates to work with the new functionality, mostly because we are no longer relying on filesystem for stable cache keys so table builders & readers need more context info to agree on cache keys. This functionality is so core, a huge number of existing tests exercise the cache key functionality. ### Performance Create db with `TEST_TMPDIR=/dev/shm ./db_bench -bloom_bits=10 -benchmarks=fillrandom -num=3000000 -partition_index_and_filters` And test performance with `TEST_TMPDIR=/dev/shm ./db_bench -readonly -use_existing_db -bloom_bits=10 -benchmarks=readrandom -num=3000000 -duration=30 -cache_index_and_filter_blocks -cache_size=250000 -threads=4` using DEBUG_LEVEL=0 and simultaneous before & after runs. Before ops/sec, avg over 100 runs: 121924 After ops/sec, avg over 100 runs: 125385 (+2.8%) ### Collision probability I have built a tool, ./cache_bench -stress_cache_key to broadly simulate host-wide cache activity over many months, by making some pessimistic simplifying assumptions: * Every generated file has a cache entry for every byte offset in the file (contiguous range of cache keys) * All of every file is cached for its entire lifetime We use a simple table with skewed address assignment and replacement on address collision to simulate files coming & going, with quite a variance (super-Poisson) in ages. Some output with `./cache_bench -stress_cache_key -sck_keep_bits=40`: ``` Total cache or DBs size: 32TiB Writing 925.926 MiB/s or 76.2939TiB/day Multiply by 9.22337e+18 to correct for simulation losses (but still assume whole file cached) ``` These come from default settings of 2.5M files per day of 32 MB each, and `-sck_keep_bits=40` means that to represent a single file, we are only keeping 40 bits of the 128-bit cache key. With file size of 2\*\*25 contiguous keys (pessimistic), our simulation is about 2\*\*(128-40-25) or about 9 billion billion times more prone to collision than reality. More default assumptions, relatively pessimistic: * 100 DBs in same process (doesn't matter much) * Re-open DB in same process (new session ID related to old session ID) on average every 100 files generated * Restart process (all new session IDs unrelated to old) 24 times per day After enough data, we get a result at the end: ``` (keep 40 bits) 17 collisions after 2 x 90 days, est 10.5882 days between (9.76592e+19 corrected) ``` If we believe the (pessimistic) simulation and the mathematical generalization, we would need to run a billion machines all for 97 billion days to expect a cache key collision. To help verify that our generalization ("corrected") is robust, we can make our simulation more precise with `-sck_keep_bits=41` and `42`, which takes more running time to get enough data: ``` (keep 41 bits) 16 collisions after 4 x 90 days, est 22.5 days between (1.03763e+20 corrected) (keep 42 bits) 19 collisions after 10 x 90 days, est 47.3684 days between (1.09224e+20 corrected) ``` The generalized prediction still holds. With the `-sck_randomize` option, we can see that we are beating "random" cache keys (except offsets still non-randomized) by a modest amount (roughly 20x less collision prone than random), which should make us reasonably comfortable even in "degenerate" cases: ``` 197 collisions after 1 x 90 days, est 0.456853 days between (4.21372e+18 corrected) ``` I've run other tests to validate other conditions behave as expected, never behaving "worse than random" unless we start chopping off structured data. Reviewed By: zhichao-cao Differential Revision: D33171746 Pulled By: pdillinger fbshipit-source-id: f16a57e369ed37be5e7e33525ace848d0537c88f
2021-12-17 01:13:55 +00:00
// Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
#include "cache/cache_key.h"
#include <algorithm>
#include <atomic>
#include "rocksdb/cache.h"
#include "table/unique_id_impl.h"
#include "util/hash.h"
#include "util/math.h"
namespace ROCKSDB_NAMESPACE {
// Value space plan for CacheKey:
//
// session_etc64_ | offset_etc64_ | Only generated by
// ---------------+---------------+------------------------------------------
// 0 | 0 | Reserved for "empty" CacheKey()
// 0 | > 0, < 1<<63 | CreateUniqueForCacheLifetime
// 0 | >= 1<<63 | CreateUniqueForProcessLifetime
// > 0 | any | OffsetableCacheKey.WithOffset
CacheKey CacheKey::CreateUniqueForCacheLifetime(Cache *cache) {
// +1 so that we can reserve all zeros for "unset" cache key
uint64_t id = cache->NewId() + 1;
// Ensure we don't collide with CreateUniqueForProcessLifetime
assert((id >> 63) == 0U);
return CacheKey(0, id);
}
CacheKey CacheKey::CreateUniqueForProcessLifetime() {
// To avoid colliding with CreateUniqueForCacheLifetime, assuming
// Cache::NewId counts up from zero, here we count down from UINT64_MAX.
// If this ever becomes a point of contention, we could use CoreLocalArray.
static std::atomic<uint64_t> counter{UINT64_MAX};
uint64_t id = counter.fetch_sub(1, std::memory_order_relaxed);
// Ensure we don't collide with CreateUniqueForCacheLifetime
assert((id >> 63) == 1U);
return CacheKey(0, id);
}
// Value plan for CacheKeys from OffsetableCacheKey, assuming that
// db_session_ids are generated from a base_session_id and
// session_id_counter (by SemiStructuredUniqueIdGen+EncodeSessionId
// in DBImpl::GenerateDbSessionId):
//
// Conceptual inputs:
// db_id (unstructured, from GenerateRawUniqueId or equiv)
// * could be shared between cloned DBs but rare
// * could be constant, if session id suffices
// base_session_id (unstructured, from GenerateRawUniqueId)
// session_id_counter (structured)
// * usually much smaller than 2**24
// file_number (structured)
// * usually smaller than 2**24
// offset_in_file (structured, might skip lots of values)
// * usually smaller than 2**32
// max_offset determines placement of file_number to prevent
// overlapping with offset
//
// Outputs come from bitwise-xor of the constituent pieces, low bits on left:
//
// |------------------------- session_etc64 -------------------------|
// | +++++++++++++++ base_session_id (lower 64 bits) +++++++++++++++ |
// |-----------------------------------------------------------------|
// | session_id_counter ...| |
// |-----------------------------------------------------------------|
// | | ... file_number |
// | | overflow & meta |
// |-----------------------------------------------------------------|
//
//
// |------------------------- offset_etc64 --------------------------|
// | hash of: ++++++++++++++++++++++++++++++++++++++++++++++++++++++ |
// | * base_session_id (upper ~39 bits) |
// | * db_id (~122 bits entropy) |
// |-----------------------------------------------------------------|
// | offset_in_file ............... | |
// |-----------------------------------------------------------------|
// | | file_number, 0-3 |
// | | lower bytes |
// |-----------------------------------------------------------------|
//
// Based on max_offset, a maximal number of bytes 0..3 is chosen for
// including from lower bits of file_number in offset_etc64. The choice
// is encoded in two bits of metadata going into session_etc64, though
// the common case of 3 bytes is encoded as 0 so that session_etc64
// is unmodified by file_number concerns in the common case.
//
// There is nothing preventing "file number overflow & meta" from meeting
// and overlapping with session_id_counter, but reaching such a case requires
// an intractable combination of large file offsets (thus at least some large
// files), large file numbers (thus large number of files generated), and
// large number of session IDs generated in a single process. A trillion each
// (2**40) of session ids, offsets, and file numbers comes to 120 bits.
// With two bits of metadata and byte granularity, this is on the verge of
// overlap, but even in the overlap case, it doesn't seem likely that
// a file from billions of files or session ids ago will still be live
// or cached.
//
// In fact, if our SST files are all < 4TB (see
// BlockBasedTable::kMaxFileSizeStandardEncoding), then SST files generated
// in a single process are guaranteed to have unique cache keys, unless/until
// number session ids * max file number = 2**86, e.g. 1 trillion DB::Open in
// a single process and 64 trillion files generated. Even at that point, to
// see a collision we would need a miraculous re-synchronization of session
// id and file number, along with a live file or stale cache entry from
// trillions of files ago.
//
// How https://github.com/pdillinger/unique_id applies here:
// Every bit of output always includes "unstructured" uniqueness bits and
// often combines with "structured" uniqueness bits. The "unstructured" bits
// change infrequently: only when we cannot guarantee our state tracking for
// "structured" uniqueness hasn't been cloned. Using a static
// SemiStructuredUniqueIdGen for db_session_ids, this means we only get an
// "all new" session id when a new process uses RocksDB. (Between processes,
// we don't know if a DB or other persistent storage has been cloned.) Within
// a process, only the session_lower of the db_session_id changes
// incrementally ("structured" uniqueness).
//
// This basically means that our offsets, counters and file numbers allow us
// to do somewhat "better than random" (birthday paradox) while in the
// degenerate case of completely new session for each tiny file, we still
// have strong uniqueness properties from the birthday paradox, with ~103
// bit session IDs or up to 128 bits entropy with different DB IDs sharing a
// cache.
//
// More collision probability analysis:
// Suppose a RocksDB host generates (generously) 2 GB/s (10TB data, 17 DWPD)
// with average process/session lifetime of (pessimistically) 4 minutes.
// In 180 days (generous allowable data lifespan), we generate 31 million GB
// of data, or 2^55 bytes, and 2^16 "all new" session IDs.
//
// First, suppose this is in a single DB (lifetime 180 days):
// 128 bits cache key size
// - 55 <- ideal size for byte offsets + file numbers
// - 2 <- bits for offsets and file numbers not exactly powers of two
// - 2 <- bits for file number encoding metadata
// + 2 <- bits saved not using byte offsets in BlockBasedTable::GetCacheKey
// ----
// 71 <- bits remaining for distinguishing session IDs
// The probability of a collision in 71 bits of session ID data is less than
// 1 in 2**(71 - (2 * 16)), or roughly 1 in a trillion. And this assumes all
// data from the last 180 days is in cache for potential collision, and that
// cache keys under each session id exhaustively cover the remaining 57 bits
// while in reality they'll only cover a small fraction of it.
//
// Although data could be transferred between hosts, each host has its own
// cache and we are already assuming a high rate of "all new" session ids.
// So this doesn't really change the collision calculation. Across a fleet
// of 1 million, each with <1 in a trillion collision possibility,
// fleetwide collision probability is <1 in a million.
//
// Now suppose we have many DBs per host, say 2**10, with same host-wide write
// rate and process/session lifetime. File numbers will be ~10 bits smaller
// and we will have 2**10 times as many session IDs because of simultaneous
// lifetimes. So now collision chance is less than 1 in 2**(81 - (2 * 26)),
// or roughly 1 in a billion.
//
// Suppose instead we generated random or hashed cache keys for each
// (compressed) block. For 1KB compressed block size, that is 2^45 cache keys
// in 180 days. Collision probability is more easily estimated at roughly
// 1 in 2**(128 - (2 * 45)) or roughly 1 in a trillion (assuming all
// data from the last 180 days is in cache, but NOT the other assumption
// for the 1 in a trillion estimate above).
//
// Conclusion: Burning through session IDs, particularly "all new" IDs that
// only arise when a new process is started, is the only way to have a
// plausible chance of cache key collision. When processes live for hours
// or days, the chance of a cache key collision seems more plausibly due
// to bad hardware than to bad luck in random session ID data.
//
OffsetableCacheKey::OffsetableCacheKey(const std::string &db_id,
const std::string &db_session_id,
uint64_t file_number,
uint64_t max_offset) {
#ifndef NDEBUG
max_offset_ = max_offset;
#endif
// Closely related to GetSstInternalUniqueId, but only need 128 bits and
// need to include an offset within the file.
// See also https://github.com/pdillinger/unique_id for background.
uint64_t session_upper = 0; // Assignment to appease clang-analyze
uint64_t session_lower = 0; // Assignment to appease clang-analyze
{
Status s = DecodeSessionId(db_session_id, &session_upper, &session_lower);
if (!s.ok()) {
// A reasonable fallback in case malformed
Hash2x64(db_session_id.data(), db_session_id.size(), &session_upper,
&session_lower);
}
}
// Hash the session upper (~39 bits entropy) and DB id (120+ bits entropy)
// for more global uniqueness entropy.
// (It is possible that many DBs descended from one common DB id are copied
// around and proliferate, in which case session id is critical, but it is
// more common for different DBs to have different DB ids.)
uint64_t db_hash = Hash64(db_id.data(), db_id.size(), session_upper);
// This establishes the db+session id part of the cache key.
//
// Exactly preserve (in common cases; see modifiers below) session lower to
// ensure that session ids generated during the same process lifetime are
// guaranteed unique.
//
// We put this first for CommonPrefixSlice(), so that a small-ish set of
// cache key prefixes to cover entries relevant to any DB.
session_etc64_ = session_lower;
// This provides extra entopy in case of different DB id or process
// generating a session id, but is also partly/variably obscured by
// file_number and offset (see below).
offset_etc64_ = db_hash;
// Into offset_etc64_ we are (eventually) going to pack & xor in an offset and
// a file_number, but we might need the file_number to overflow into
// session_etc64_. (There must only be one session_etc64_ value per
// file, and preferably shared among many files.)
//
// Figure out how many bytes of file_number we are going to be able to
// pack in with max_offset, though our encoding will only support packing
// in up to 3 bytes of file_number. (16M file numbers is enough for a new
// file number every second for half a year.)
int file_number_bytes_in_offset_etc =
(63 - FloorLog2(max_offset | 0x100000000U)) / 8;
int file_number_bits_in_offset_etc = file_number_bytes_in_offset_etc * 8;
// Assert two bits of metadata
assert(file_number_bytes_in_offset_etc >= 0 &&
file_number_bytes_in_offset_etc <= 3);
// Assert we couldn't have used a larger allowed number of bytes (shift
// would chop off bytes).
assert(file_number_bytes_in_offset_etc == 3 ||
(max_offset << (file_number_bits_in_offset_etc + 8) >>
(file_number_bits_in_offset_etc + 8)) != max_offset);
uint64_t mask = (uint64_t{1} << (file_number_bits_in_offset_etc)) - 1;
// Pack into high bits of etc so that offset can go in low bits of etc
// TODO: could be EndianSwapValue?
uint64_t offset_etc_modifier = ReverseBits(file_number & mask);
assert(offset_etc_modifier << file_number_bits_in_offset_etc == 0U);
// Overflow and 3 - byte count (likely both zero) go into session_id part
uint64_t session_etc_modifier =
(file_number >> file_number_bits_in_offset_etc << 2) |
static_cast<uint64_t>(3 - file_number_bytes_in_offset_etc);
// Packed into high bits to minimize interference with session id counter.
session_etc_modifier = ReverseBits(session_etc_modifier);
// Assert session_id part is only modified in extreme cases
assert(session_etc_modifier == 0 || file_number > /*3 bytes*/ 0xffffffU ||
max_offset > /*5 bytes*/ 0xffffffffffU);
// Xor in the modifiers
session_etc64_ ^= session_etc_modifier;
offset_etc64_ ^= offset_etc_modifier;
// Although DBImpl guarantees (in recent versions) that session_lower is not
// zero, that's not entirely sufficient to guarantee that session_etc64_ is
// not zero (so that the 0 case can be used by CacheKey::CreateUnique*)
if (session_etc64_ == 0U) {
session_etc64_ = session_upper | 1U;
}
assert(session_etc64_ != 0);
}
} // namespace ROCKSDB_NAMESPACE