FastCache.Cached 1.5.0

FastCache.Cached

<p><img src="https://raw.githubusercontent.com/neon-sunset/fast-cache/main/img/cached-small-transparent.png" width="180" height="180" align="right" /></p>

The fastest cache library written in C# for items with set expiration time. Easy to use, thread-safe and light on memory.

Optimized to scale from dozens to millions of items. Features lock-free reads and writes, allocation-free reads and automatic eviction.

Credit to Vladimir Sadov for his implementation of NonBlocking.ConcurrentDictionary which is used as an underlying store.

Quick start

dotnet add package FastCache.Cached or Install-Package FastCache.Cached

Get cached value or save a new one with expiration of 60 minutes

public FinancialReport GetReport(int month, int year)
{
  if (Cached<FinancialReport>.TryGet(month, year, out var cached))
  {
    return cached.Value;
  }

  var report = // Expensive computation: retrieve data and calculate report

  return cached.Save(report, TimeSpan.FromMinutes(60));
}

Wrap and cache the result of a regular method call

var report = Cached.GetOrCompute(month, year, GetReport, TimeSpan.FromMinutes(60));

Or an async one

// For methods that return Task<T> or ValueTask<T>
var report = await Cached.GetOrCompute(month, year, GetReportAsync, TimeSpan.FromMinutes(60));

Save the value to cache but only if the cache size is below limit

public FinancialReport GetReport(int month, int year)
{
  if (Cached<FinancialReport>.TryGet(month, year, out var cached))
  {
    return cached.Value;
  }

  return cached.Save(report, TimeSpan.FromMinutes(60), limit: 2_500_000);
}

// GetOrCompute with maximum cache size limit.
// RAM is usually plenty but what if the user runs Chrome?
var report = Cached.GetOrCompute(month, year, GetReport, TimeSpan.FromMinutes(60), limit: 2_500_000);

Add new data without accessing cache item first (e.g. loading a large batch of independent values to cache)

using FastCache.Extensions;
...
foreach (var ((month, year), report) in reportsResultBatch)
{
  report.Cache(month, year, TimeSpan.FromMinutes(60));
}

Store common type (string) in a shared cache store (other users may share the cache for the same parameter type, this time it's int)

// GetOrCompute<...V> where V is string.
// To save some other string for the same 'int' number simultaneously, look at the option below :)
var userNote = Cached.GetOrCompute(userId, GetUserNoteString, TimeSpan.FromMinutes(5));

Or in a separate one by using value object (Recommended)

readonly record struct UserNote(string Value);

// GetOrCompute<...V> where V is UserNote
var userNote = Cached.GetOrCompute(userId, GetUserNote, TimeSpan.FromMinutes(5));

// This is how it looks for TryGet
if (Cached<UserNote>.TryGet(userId, out var cached))
{
  return cached.Value;
}
...
return cached.Save(userNote, TimeSpan.FromMinutes(5));

Features and design philosophy

• In-memory cache for items with expiration time and automatic eviction
• Little to no ceremony - no need to configure or initialize, just add the package and you are ready to go. Behavior can be further customized via env variables
• Focused design allows to reduce memory footprint per item and minimize overhead via inlining and static dispatch
• High performance and scaling covering both simplest applications and highly loaded services. Can handle 1-100M+ items with O(1) read/write time and up to O(n~) memory cost/cpu time cost for full eviction
• Lock-free and wait-free reads/writes of cached items. Performance will improve with threads, data synchronization cost is minimal thanks to NonBlocking.ConcurrentDictionary
• Multi-key store access without collisions between key types. Collisions are avoided by statically dispatching on the composite key type signature e.g. (string, CustomEnum, int) together with the type of cached value - composite keys are structurally evaluated for equality, different combinations will correspond to different cache items
• Handles timezone/DST updates on most platforms by relying on system uptime timestamp for item expiration - Environment.TickCount64 which is also significantly faster than DateTime.UtcNow

Performance

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.22000
AMD Ryzen 7 5800X, 1 CPU, 16 logical and 8 physical cores
.NET 6.0.5 (6.0.522.21309), X64 RyuJIT

TLDR: FastCache.Cached vs Microsoft.Extensions.Caching.Memory.MemoryCache

+CachedRange.Save(ReadOnlySpan<(K, V)>) provides parallelized bulk writes out of box

++CacheManager doesn't have read throughput results because test suite would take too long to run to include CacheManager and LazyCache. Given higher CPU usage by CacheManager and higher RAM usage by LazyCache it is reasonable to assume they would score lower and scale worse due to higher number of locks

Read/Write lowest achievable latency

Read throughput detailed

Further reading

• Keys and composite keys performance estimation: Code / Results

Notes

• FastCache.Cached defaults provide highest performance and don't require spending time on finding a way to use API optimally. The design goal is to nudge a developer to use the fastest way to achieve his or her goals while strictly adhering to the principle of "pay for play".
• Comparison was made with a string-based key. Composite keys supported by FastCache.Cached have additional performance cost. Numeric keys have significantly higher performance in some instances (down to 5ns read latency on x86_64 where K is int or uint and V fits in a single register).
• *CacheManger documentation suggests using WithMicrosoftMemoryCacheHandle() by default which has terrible performance (and uses obsolete System.Runtime.Caching). We give it a better fighting chance by using WithDictionaryHandle() instead.
• **CachedRange.Save(ReadOnlyMemory<(K, V)>) (and CachedRange.Save(ReadOnlyMemory<K>, ReadOnlyMemory<V>)) automatically splits the range into slices of 1024 or length / cpu cores elements and saves them in parallel which dramatically improves throughput. This works especially well because the execution flow is lock-free and does not suffer from false sharing of CPU cache lines (except for the adjacent entries at the edges of slices)
• ***Forced single-threaded execution of CachedRange.Save(ReadOnlyMemory<(K, V)>)
• Overall, performance stays relatively comparable when downgrading to .NET 5 and decreases further by 15-30% when using .NET Core 3.1 with the difference ratio between libraries staying close to provided above.
• Non-standard platforms (the ones that aren't CLR based) use DateTime.UtcNow fallback instead of Environment.TickCount64, which will perform slower depending on the platform-specific implementation.

On benchmark data

Throughput saturation means that all necessary data structures are fully available in the CPU cache and branch predictor has learned branch patters of the executed code. This is only possible in scenarios such as items being retrieved or added/updated in a tight loop or very frequently on the same cores. This means that real world performance will not saturate maximum throughput and will be bottlenecked by memory access latency and branch misprediction stalls. As a result, you can expect resulting performance variance of 1-10x min latency depending on hardware and outside factors.

From 🇺🇦 Ukraine with ♥️