Messages in this thread | | | Date | Wed, 04 Sep 2013 22:35:40 -0400 | From | Waiman Long <> | Subject | Re: [PATCH v7 1/4] spinlock: A new lockref structure for lockless update of refcount |
| |
On 09/04/2013 05:34 PM, Linus Torvalds wrote: > On Wed, Sep 4, 2013 at 12:25 PM, Waiman Long<waiman.long@hp.com> wrote: >> Yes, the perf profile was taking from an 80-core machine. There isn't any >> scalability issue hiding for the short workload on an 80-core machine. >> >> However, I am certain that more may pop up when running in an even larger >> machine like the prototype 240-core machine that our team has been testing >> on. > Sure. Please let us know, I think it's going to be interesting to see > what that shows. > > SGI certainly did much larger machines, but their primary target > tended to be all user space, so they had things like "tons of > concurrent page faults in the same process" rather than filename > lookup or the tty layer. > > Linus
I think SGI is more focus on compute-intensive workload. HP is more focus on high-end commercial workload like SAP HANA. Below was a sample perf profile of the high-systime workload on a 240-core prototype machine (HT off) with 3.10-rc1 kernel with my lockref and seqlock patches:
9.61% 3382925 swapper [kernel.kallsyms] [k] _raw_spin_lock |--59.90%-- rcu_process_callbacks |--19.41%-- load_balance |--9.58%-- rcu_accelerate_cbs |--6.70%-- tick_do_update_jiffies64 |--1.46%-- scheduler_tick |--1.17%-- sched_rt_period_timer |--0.56%-- perf_adjust_freq_unthr_context --1.21%-- [...]
6.34% 99 reaim [kernel.kallsyms] [k] _raw_spin_lock |--73.96%-- load_balance |--11.98%-- rcu_process_callbacks |--2.21%-- __mutex_lock_slowpath |--2.02%-- rcu_accelerate_cbs |--1.95%-- wake_up_new_task |--1.70%-- scheduler_tick |--1.67%-- xfs_alloc_log_agf |--1.24%-- task_rq_lock |--1.15%-- try_to_wake_up --2.12%-- [...]
5.39% 2 reaim [kernel.kallsyms] [k] _raw_spin_lock_irqsave |--95.08%-- rwsem_wake |--1.80%-- rcu_process_callbacks |--1.03%-- prepare_to_wait |--0.59%-- __wake_up --1.50%-- [...]
2.28% 1 reaim [kernel.kallsyms] [k] _raw_spin_lock_irq |--90.56%-- rwsem_down_write_failed |--9.25%-- __schedule --0.19%-- [...]
Longman
| |