Skip to main content

Computer Science: Arbitrarily Fine Locking

This is a relatively simple idea concerning mutex usage. I imagine someone else has probably thought of it before. However, since I just thought of it, I figured I'd blog it. I have no clue why I was thinking about mutexes. I usually prefer share-nothing approaches like Erlang. Note, I am specifically not trying to comment on Python's GIL.

Imagine you have a bunch of resources (like objects or structs) that you wish to protect. One way to protect them is to use a single mutex. This is called coarse-grained locking. At the opposite end of the spectrum, you can create a new mutex for every resource. This is called fine-grained locking. However, what if you want something in the middle?

Having a single lock is unfortunate because it forces a lot of your code to be effectively single-threaded, even if you have multiple processors. However, perhaps creating a new mutex for every single resource might be overkill. (Personally, I'm unsure of why that might be the case.)

Here's an approach to get arbitrarily fine locking. Create N mutexes (where N is tunable). Protect each resource using mutex number resource_id % N. The resource_id could be whatever, as long as it's unique. Perhaps it's the index of an array, or perhaps it's a pointer to the resource in memory.

And now for something completely different! The best part of Lisp is that it has garbage collection. It recycles garbage so that you can grow new trees ;)

Comments

Benjamin said…
Your assigning a resource to a lock by hashing in the hopes that you'll have a large pool of resources more or less randomly distributed across a smaller pool of locks. Clever, but I'm still not sure what problem this solves.

Don't forget: you've got to assure consistent ordering of lock acquisition. Otherwise you're setting yourself up for deadlocks if you have tasks that require more than one resource at a time.

Also: acquiring and releasing locks is not free. This will place an upper bound on how many you'd want to use in your system.
> Clever, but I'm still not sure what problem this solves.

I'm not either. I'm not sure why it even came to mind.

> Don't forget: you've got to assure consistent ordering of lock acquisition. Otherwise you're setting yourself up for deadlocks if you have tasks that require more than one resource at a time.

Yep.

> Also: acquiring and releasing locks is not free. This will place an upper bound on how many you'd want to use in your system.

Yep.
Bill Mill said…
> The resource_id could be whatever, as long as it's unique

It needs to be unique and evenly distributed with respect to mod N, not just unique.
> It needs to be unique and evenly distributed with respect to mod N, not just unique.

Agreed, although you can pull the same tricks that you pull with hashes. If you get too many collisions on any one mutex, you can increase N.
Anonymous said…
It is not a bad idea, but it is not new either. We have been doing this a lot when we implement hash tables for dynamic hashing (linked lists growing out of buckets). You have a lock per bucket, so that insert and delete operations can work atomically. But if you have many buckets and many cores/threads, many of those operations can take place at the same time, just not on the same bucket.

So, yeah, it works and has the desired effect, but it has been done many times before.
Kris said…
This is used in the FreeBSD kernel in some places where a mutex is needed but either allocating one statically or at runtime would have too much time or space overhead.

We keep a "pool" of mutexes that can be used by anyone and are hashed based on a resource address. To get around the deadlock issue pool mutexes must be leaf mutexes, i.e. you are not allowed to acquire other locks while holding them.
> It is not a bad idea, but it is not new either.

> This is used in the FreeBSD kernel in some places

Excellent comments. Thanks!

Ideas are rarely new, so I'm always glad to hear when I had a good, existing idea rather than a bad, existing idea ;)