From: Matthew Wilcox > Sent: 07 March 2024 19:55 > > I've had a few conversations recently about how many objects should be > in a batch in some disparate contextx, so I thought I'd write down my > opinion so I can refer to it in future. TLDR: Start your batch size > around 10, adjust the batch size when measurements tell you to change it. > > In this model, let's look at the cost of allocating N objects from an > allocator. Assume there's a fixed cost, say 4 (units are not relevant > here) for going into the allocator and then there's a 1 unit cost per > object (eg we're taking a spinlock, pulling N objects out of the data > structure and releasing the spinlock again). I think you are trying to measure the length of a piece of string. (and not the ones in the box labelled 'pieces of string too small to keep') What I did recently for a global+local buffer allocator was to make each entry on the global list itself be a list of objects. So if the local list was empty it was a single cmpxchg to grab (about 100) items. Similarly if the local free list got too big a single locked operation would free a block of items. That significantly reduced lock contention. If was necessary to split the free of very long lists - otherwise a single item on the global list could contain silly numbers of items. This was userspace, and we don't talk about the test that ended up with ALL system memory on one linked list. I managed to kill enough (remote) things to get a working shell. It took the system about 20 minutes just to count the linked list! David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
