CFS Bandwidth Control
in [Kernel]
|
Prev: sata_mv driver support for Marvell Hercules II controller in newer kernels
Next: cxgb3: FIx VLAN over Jumbo frames
From: Bharata B Rao on 16 Feb 2010 00:50 On Fri, Feb 12, 2010 at 06:54:52PM -0800, Paul Turner wrote: > > The skeleton of our approach is as follows: > - As above we maintain a global pool, per-tg, pool of unassigned quota. On it > we track the bandwidth period, quota per period, and runtime remaining in > the current period. As bandwidth is used within a period it is decremented > from runtime. Runtime is currently synchronized using a spinlock, in the > current implementation there's no reason this couldn't be done using > atomic ops instead however the spinlock allows for a little more flexibility > in experimentation with other schemes. > - When a cfs_rq participating in a bandwidth constrained task_group executes > it acquires time in sysctl_sched_cfs_bandwidth_slice (default currently > 10ms) size chunks from the global pool, this synchronizes under rq->lock and > is part of the update_curr path. > - Throttled entities are dequeued immediately (as opposed to delaying this > operation to the put path), this avoids some potentially poor load-balancer > interactions and preserves the 'verbage' of the put_task semantic. > Throttled entities are gated from participating in the tree at the > {enqueue, dequeue}_entity level. They are also skipped for load > balance in the same manner as Bharatta's patch-series employs. I did defer the dequeue until next put because walking the se hierarchy multiple times (from update_curr -> dequeue_entity -> update_curr) appeared too complex when I started with it. > > Interface: > ---------- > Two new cgroupfs files are added to the cpu subsystem: > - cpu.cfs_period_us : period over which bandwidth is to be regulated > - cpu.cfs_quota_us : bandwidth available for consumption per period > > One important interface change that this introduces (versus the rate limits > proposal) is that the defined bandwidth becomes an absolute quantifier. > > e.g. a bandwidth of 5 seconds (cpu.cfs_quota_us=5000000) on a period of 1 second > (cpu.cfs_period_us=1000000) would result in 5 wall seconds of cpu time being > consumable every 1 wall second. As I have said earlier, I would like to hear what others say about this interface. Especially from Linux-vserver project since it is already using the cfs hard limit patches in their test release. Herbert ? Thanks for your work. More later when I review the individual patches in detail. Regards, Bharata. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: Bharata B Rao on 16 Feb 2010 01:20
On Fri, Feb 12, 2010 at 06:54:52PM -0800, Paul Turner wrote: > Todo: > ----- > - hierarchal nr_tasks_running accounting: > This is a deficiency currently shared with SCHED_RT rate limiting. When > entities is throttled the running tasks it owns are not subtracted from > rq->nr_running. This then results in us missing idle_balance() due to > phantom tasks and load balancer weight per task calculations being > incorrect. > > This code adds complexity which was both increasing the complexity of the > initial review for this patchset and truly probably best reviewed > independently of this feature's scope. To that end we'll post a separate > patch for this issue against the current RT rate-limiting code and merge any > converged on approach here as appropriate. I had tried updating rq->nr_running in my v2 patchset (http://lkml.org/lkml/2009/9/30/117, http://lkml.org/lkml/2009/9/30/119) But since I felt that it added a lot of complexity, I removed it subsequently in v3 (http://lkml.org/lkml/2009/11/9/65) and kept it similar to RT. > > - throttle statistics: > Some statistics regarding the frequency and duration of throttling > definitely in order. Please take a look at some of the throttling related stats I am collecting in my patchset. Regards, Bharata. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |