Parallel.ForEach vs Task.WhenAll vs PLINQ — which scales better and when

	Threads or tasks?	Best for	Throughput on I/O?	Throughput on CPU?
Parallel.ForEach	OS threads	CPU-bound batches	Wastes threads	Best
Task.WhenAll	Tasks (no extra threads on I/O)	I/O fan-out	Best	OK but no partitioning
PLINQ	OS threads + partitioning	Functional data pipelines on CPU	Wastes threads	Good (boilerplate-free)

CPU-bound — Parallel.ForEach

Parallel.ForEach(images, img => Process(img));

• Uses thread pool, partitions work, scales with Environment.ProcessorCount.
• Do not use for await inside — see "async pitfall" below.

I/O-bound — Task.WhenAll

var tasks = urls.Select(u => http.GetStringAsync(u));
var results = await Task.WhenAll(tasks);

• One thread can have thousands of in-flight I/O tasks. Do not use Parallel.ForEach here — it would create one OS thread per URL.

Throttled async fan-out — Parallel.ForEachAsync (.NET 6+)

await Parallel.ForEachAsync(urls,
    new ParallelOptions { MaxDegreeOfParallelism = 16 },
    async (u, ct) => await http.GetStringAsync(u, ct));

The sweet spot for "I have 50 000 HTTP calls and do not want to open them all at once."

Functional CPU pipelines — PLINQ

var hashes = files
    .AsParallel()
    .WithDegreeOfParallelism(8)
    .Select(f => Sha256(f))
    .ToArray();

• Cleaner than Parallel.ForEach when you are composing transformations.
• Ordering — use .AsOrdered() only if necessary, costs throughput.

The async pitfall

// Do not write this
Parallel.ForEach(urls, async u => await http.GetStringAsync(u));
// Lambda is async void — fire-and-forget; Parallel.ForEach returns before any HTTP completes.

Use Parallel.ForEachAsync instead.

Choosing matrix

Work type	Item count	Choice
CPU work, small N	4 to 8	Task.WhenAll is fine (no partitioning cost)
CPU work, large N	1000+	Parallel.ForEach or PLINQ
I/O work, any N	*	Task.WhenAll or Parallel.ForEachAsync (throttled)

Don't forget — measurement first

BenchmarkDotNet or production profiling. Parallel code often gets slower due to contention; verify your specific workload.