An Overview of Kernel Lock Improvements [pdf](events.linuxfoundation.org)

94 pointsby mmastrac12 anni fa11 comments

events.linuxfoundation.org

An Overview of Kernel Lock Improvements [pdf]

http://events.linuxfoundation.org/sites/events/files/slides/linuxcon-2014-locking-final.pdf

6 comments

xroche12 anni fa

Note: for those interested, MCS locks and qspinlocks are discussed in this fine article: http://lwn.net/Articles/590243/

MoOmer12 anni fa

This is great; I'm still reading through, but, I had no idea that the performance drop was that steep!

jebblue12 anni fa

Do you mean the drop they later used to analyze cache line coherency when adding cores? They improved this if I understood the latter results correctly.

fmstephe12 anni fa

Can anyone explain the use of 'round-robin' to describe mulit-node scenarios and 'fill-first' for single node scenarios. I initially assumed they were describing thread schedulers, but that doesn't make clear sense in these tests. Thanks in advance.

readerrrr12 anni fa

This is the key: each locker spins on local memory rather than the lock word at page 63.

yxhuvud12 anni fa

It would be interesting to see the same benchmarks for more normal amounts of cores. Most system does not have 240 cores, after all..

ck212 anni fa

You can get 20core/40thread servers for $400/mo these days:

http://www.ovh.com/us/dedicated-servers/enterprise/2014-MG-2...

or 16core/32thread for $270

http://www.ovh.com/us/dedicated-servers/enterprise/2014-MG-1...

So double socket performance drop is a realistic concern.

readerrrr12 anni fa

The latest Intel hi-end desktop cpu has 8 cores. With hyper-threading; which makes 16 threads.

I think in the next 10 years we will reach a number that is close.

MaxBarraclough12 anni fa

If you count Xeon Phi, Intel are already up to 61.

http://ark.intel.com/products/family/71840/Intel-Xeon-Phi-Co...

davidlohr12 anni fa

Most of these optimizations imply scaling by number of cores. Thus the more cores (and thus sockets/NUMA nodes), the more the benefit. Desktop-ish systems with ~4 cores don't see much gain, but nor did we introduce performance degradations.

molixiaoge12 anni fa

great