Skip to contentSkip to conversation

Load balancing behavior for sched autogroup

DISCUSSION Tim Chen 2 participants 3 messages Mar 29, 2013 – Apr 8, 2013

lore
Share & tools
lore

Tim Chen

author
#0 lore raw

During our testings of 3.8 kernel, we noticed that after the patch

Revert "sched: Update_cfs_shares at period edge" (commit 17bc14b7),

the load between the sockets or larger system can have large imbalance. For example, for a 4 socket Westmere-EX (10 cores/socket), we notice the loadings between the sockets can differ by more than a factor of 4.

We did a simple experiment that kicks off 29 simple processes that execute a tight loop. We noticed socket 3 is already starting to schedule on hyperthreaded cpus (13 loaded cpus) while socket 1 still have lots of idle cores (3 loaded cpus). Before the patch, the load was evenly distributed across sockets.
If I turn off CONFIG_SCHED_AUTOGROUP,the loads are also distributed evenly.

(load on cpus, running on 4)
socket 0 1 2 3
---------------------------------------------------------------------------------------------

cpu:    0-3               0.00                 0.00                 0.00                99.00
cpu:    4-7               0.00                 0.00                 0.00                99.20
cpu:    8-11              0.00                 0.00                 0.00                99.00
cpu:   12-15             99.20                 0.00                 0.00                 0.00
cpu:   16-19              0.00                 0.00                 0.00                99.00
cpu:   20-23              0.00                 0.00                 0.00                 0.00
cpu:   24-27              0.00                 0.00                 0.00                 0.00
cpu:   28-31              0.00                 0.00                 0.00                 0.00
cpu:   32-35             99.20                 0.00                 0.00                99.00
cpu:   36-39             99.20                99.40                99.20                 0.00
cpu:   40-43              0.00                99.40                99.40                99.20
cpu:   44-47              0.00                99.40                99.40                99.20
cpu:   48-51             99.40                 0.00                99.40                99.20
cpu:   52-55             99.20                 0.00                99.40                99.20
cpu:   56-59              0.00                 0.00                99.40                99.40
cpu:   60-63              0.00                 0.00                 0.00                99.00
cpu:   64-67              0.00                 0.00                 0.00                99.40
cpu:   68-71              0.00                 0.00                 0.00                99.40
cpu:   72-75             99.40                 0.00                 0.00                 0.00
cpu:   76-79             99.40                 0.00                 0.00                 0.00

---------------------------------------------------------------------------------------------
Loaded cpus 7 3 6 13

Is this the intended behavior of sched autogroup? I'm a bit surprised that we are reserving this much cpu bandwidth for very low load processes (or interactive processes) in other groups.

So should the sched autogroup config option be turned off by default for server system, when we are not concerned about interactivity but want to maximize throughput by balancing out the load?

Thanks for clarifying.

Tim

#1 lore raw

I'm surprised that patch would have much effect in either direction; it changes the amortization of accounting, but not the actual numbers -- especially for a persistent load. We'll take a look.

<Resend since gmail "reinterpreted" my plain-text>

Quoted 57 lines · Tim Chen

During our testings of 3.8 kernel, we noticed that after the patch

Revert "sched: Update_cfs_shares at period edge" (commit 17bc14b7),

the load between the sockets or larger system can have large imbalance. For example, for a 4 socket Westmere-EX (10 cores/socket), we notice the loadings between the sockets can differ by more than a factor of 4.

We did a simple experiment that kicks off 29 simple processes that execute a tight loop. We noticed socket 3 is already starting to schedule on hyperthreaded cpus (13 loaded cpus) while socket 1 still have lots of idle cores (3 loaded cpus). Before the patch, the load was evenly distributed across sockets.
If I turn off CONFIG_SCHED_AUTOGROUP,the loads are also distributed evenly.

(load on cpus, running on 4)
socket 0 1 2 3
---------------------------------------------------------------------------------------------

cpu:    0-3               0.00                 0.00                 0.00                99.00
cpu:    4-7               0.00                 0.00                 0.00                99.20
cpu:    8-11              0.00                 0.00                 0.00                99.00
cpu:   12-15             99.20                 0.00                 0.00                 0.00
cpu:   16-19              0.00                 0.00                 0.00                99.00
cpu:   20-23              0.00                 0.00                 0.00                 0.00
cpu:   24-27              0.00                 0.00                 0.00                 0.00
cpu:   28-31              0.00                 0.00                 0.00                 0.00
cpu:   32-35             99.20                 0.00                 0.00                99.00
cpu:   36-39             99.20                99.40                99.20                 0.00
cpu:   40-43              0.00                99.40                99.40                99.20
cpu:   44-47              0.00                99.40                99.40                99.20
cpu:   48-51             99.40                 0.00                99.40                99.20
cpu:   52-55             99.20                 0.00                99.40                99.20
cpu:   56-59              0.00                 0.00                99.40                99.40
cpu:   60-63              0.00                 0.00                 0.00                99.00
cpu:   64-67              0.00                 0.00                 0.00                99.40
cpu:   68-71              0.00                 0.00                 0.00                99.40
cpu:   72-75             99.40                 0.00                 0.00                 0.00
cpu:   76-79             99.40                 0.00                 0.00                 0.00

---------------------------------------------------------------------------------------------
Loaded cpus 7 3 6 13

Is this the intended behavior of sched autogroup? I'm a bit surprised that we are reserving this much cpu bandwidth for very low load processes (or interactive processes) in other groups.

So should the sched autogroup config option be turned off by default for server system, when we are not concerned about interactivity but want to maximize throughput by balancing out the load?

Thanks for clarifying.

Tim

author
#2 lore raw
Quoted 4 lines · Paul Turner

I'm surprised that patch would have much effect in either direction; it changes the amortization of accounting, but not the actual numbers -- especially for a persistent load. We'll take a look.

Hi Paul,

Wonder if you have a chance to take a look at the load imbalance with sched autogroup?

Thanks.

Tim