5 ms·

An Overview of Kernel Lock Improvements [pdf]

94 points by mmastrac 12y ago

MoOmer 12y ago
This is great; I'm still reading through, but, I had no idea that the performance drop was that steep!
- jebblue 12y ago
  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.
molixiaoge 12y ago
great
xroche 12y ago
Note: for those interested, MCS locks and qspinlocks are discussed in this fine article: http://lwn.net/Articles/590243/ http://lwn.net/Articles/590243/
readerrrr 12y ago
This is the key: each locker spins on local memory rather than the lock word at page 63.
yxhuvud 12y ago
It would be interesting to see the same benchmarks for more normal amounts of cores. Most system does not have 240 cores, after all..
- readerrrr 12y ago
  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.
  - MaxBarraclough 12y ago
    If you count Xeon Phi, Intel are already up to 61. http://ark.intel.com/products/family/71840/Intel-Xeon-Phi-Coprocessors http://ark.intel.com/products/family/71840/Intel-Xeon-Phi-Co...
- ck2 12y ago
  You can get 20core/40thread servers for $400/mo these days: http://www.ovh.com/us/dedicated-servers/enterprise/2014-MG-256.xml 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-128.xml http://www.ovh.com/us/dedicated-servers/enterprise/2014-MG-1... So double socket performance drop is a realistic concern.
- davidlohr 12y ago
  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.
fmstephe 12y ago
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.

94 points by mmastrac 12y ago

MoOmer 12y ago

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

jebblue 12y ago
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.

molixiaoge 12y ago

great

xroche 12y ago

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

readerrrr 12y ago

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

yxhuvud 12y ago

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

readerrrr 12y ago
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.
- MaxBarraclough 12y ago
  If you count Xeon Phi, Intel are already up to 61. http://ark.intel.com/products/family/71840/Intel-Xeon-Phi-Coprocessors http://ark.intel.com/products/family/71840/Intel-Xeon-Phi-Co...
ck2 12y ago
You can get 20core/40thread servers for $400/mo these days: http://www.ovh.com/us/dedicated-servers/enterprise/2014-MG-256.xml 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-128.xml http://www.ovh.com/us/dedicated-servers/enterprise/2014-MG-1... So double socket performance drop is a realistic concern.
davidlohr 12y ago
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.

fmstephe 12y ago

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.