Andrea Righi wrote:
> I'm resending an updated version of the cgroup I/O bandwidth controller
> that I wrote some months ago (http://lwn.net/Articles/265944), since
> someone asked me recently.
> 
> The goal of this patchset is to implement a block device I/O bandwidth
> controller using cgroups.
> 
> Detailed informations about design, its goal and usage are described in
> the documentation.
> 
> Review, comments and feedbacks are welcome.
> 

Hi, Andrea,

There are several parallel efforts for the IO controller? Could you
describe/compare them with yours? Is any consensus building taking place?

CC'ing containers mailing list.

-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL
_______________________________________________
Containers mailing list
Containers@lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers

Balbir Singh wrote:
> Andrea Righi wrote:
>> I'm resending an updated version of the cgroup I/O bandwidth controller
>> that I wrote some months ago (http://lwn.net/Articles/265944), since
>> someone asked me recently.
>>
>> The goal of this patchset is to implement a block device I/O bandwidth
>> controller using cgroups.
>>
>> Detailed informations about design, its goal and usage are described in
>> the documentation.
>>
>> Review, comments and feedbacks are welcome.
>>
> 
> Hi, Andrea,
> 
> There are several parallel efforts for the IO controller? Could you
> describe/compare them with yours? Is any consensus building taking place?
> 
> CC'ing containers mailing list.
> 

Hi Balbir,

I've seen the Ryo Tsuruta's dm-band (http://lwn.net/Articles/266257),
the CFQ cgroup solution proposed by Vasily Tarasov
(http://lwn.net/Articles/274652) and a similar approach by Satoshi
Uchida (http://lwn.net/Articles/275944).

First one is implemented at the device mapper layer and AFAIK at the
moment it allows only to define per-task, per-user and per-group rules
(cgroup support is in the TODO list anyway).

Second and third solutions are implemented at the i/o scheduler
layer, CFQ in particular.

They work as expected with direct i/o operations (or synchronous reads).
The problem is: how to limit the i/o activity of an applicaton that
already wrote all the data in memory? The i/o scheduler is not the right
place to throttle application writes, because it's "too late". At the
very least we can map back the i/o path to find which process originally
dirtied memory and depending on this information delay the dispatching
of the requests. However, this needs bigger buffers, more page cache
usage, that can lead to potential OOM conditions in massively shared
environments.

So, my approach has the disadvantage or the advantage (depending on the
context and the requirements) to explicitly choke applications'
requests. Other solutions that operates in the subsystem used to
dispatch i/o requests are probably better to maximize overall
performance, but do not offer the same control over a real QoS as
request limiting can do.

Another difference is that all of them are priority/weighted based. The
io-throttle controller, instead, allows to define direct bandwidth
limiting rules. The difference here is that priority based approach
propagates bursts and does no smoothing. Bandwidth limiting approach
controls bursts by smoothing the i/o rate. This means better performance
predictability at the cost of poor throughput optimization.

I'll run some benchmarks and post the results ASAP. It would be also
interesting to run the same benchmarks using the other i/o controllers
and compare the results in terms of fairness, performance
predictability, responsiveness, throughput, etc. I'll see what I can do.

-Andrea
_______________________________________________
Containers mailing list
Containers@lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers

Andrea Righi wrote:
> So, my approach has the disadvantage or the advantage (depending on the
> context and the requirements) to explicitly choke applications'
> requests. Other solutions that operates in the subsystem used to
> dispatch i/o requests are probably better to maximize overall
> performance, but do not offer the same control over a real QoS as
> request limiting can do.
> 
> Another difference is that all of them are priority/weighted based. The
> io-throttle controller, instead, allows to define direct bandwidth
> limiting rules. The difference here is that priority based approach
> propagates bursts and does no smoothing. Bandwidth limiting approach
> controls bursts by smoothing the i/o rate. This means better performance
> predictability at the cost of poor throughput optimization.
> 
> I'll run some benchmarks and post the results ASAP. It would be also
> interesting to run the same benchmarks using the other i/o controllers
> and compare the results in terms of fairness, performance
> predictability, responsiveness, throughput, etc. I'll see what I can do.

Here is the test report of the benchmarks. Comments, suggestions,
disapprovals, etc. are welcome.

Tests are mainly focused to demonstrate how the io-throttle controller
can be used to improve the i/o performance predictability on shared
block devices (usually any block device shared by many users and
processes, that can be grouped togheter by cgroup).

The goal of the tests is *not* to compare the overall performance of the
different solutions.

Tests have been performed using an ext3 fs + CFQ i/o scheduler on
/dev/sda.

Following some details of the testbed host:
# hdparm -i /dev/sda
/dev/sda:

 Model=SAMSUNG HS08XJC , FwRev=GR100-01,
 Config={ HardSect NotMFM HdSw>15uSec Fixed DTR>10Mbs }
 RawCHS=16383/16/63, TrkSize=34902, SectSize=0, ECCbytes=4
 BuffType=DualPortCache, BuffSize=7986kB, MaxMultSect=16, MultSect=?8?
 CurCHS=16383/16/63, CurSects=16514064, LBA=yes, LBAsects=156301488
 IORDY=on/off, tPIO={min:120,w/IORDY:120}, tDMA={min:120,rec:120}
 PIO modes:  pio0 pio1 pio2 pio3 pio4
 DMA modes:  mdma0 mdma1 mdma2
 UDMA modes: udma0 udma1 udma2 udma3 udma4 *udma5
 AdvancedPM=yes: unknown setting WriteCache=enabled
 Drive conforms to: ATA/ATAPI-6 T13 1410D revision 1:  ATA/ATAPI-1,2,3,4,5,6,7

 * signifies the current active mode

# grep "model name" /proc/cpuinfo
model name	: Intel(R) Core(TM)2 CPU         U7600  @ 1.20GHz
model name	: Intel(R) Core(TM)2 CPU         U7600  @ 1.20GHz

# grep MemTotal /proc/meminfo
MemTotal:      2021616 kB

Latest io-throttle patch, raw data of the benchmarks and scripts used to
calculate the results can be found here:

http://download.systemimager.org/~arighi/linux/patches/io-throttle/

== Test #0: basic i/o ops ==

The goal of this pre-test is to demonstrate the basic functionalities of
the io-throttle controller. In particular to show that the i/o bandwidth
limiting is applied only on the actual i/o and does not involve
read/write operations in the page cache.

A simple script (dd-band.sh) creates a 64MiB file, re-write it (in the
page cache), read the file in O_DIRECT mode and re-read the file from
the page cache.

The i/o limitations are performed during the first write (because it
generates real writes processed asynchronously by the pdflush kernel
thread outside the context of the process) and during the O_DIRECT read.
Re-writes and re-reads that only affect the page cache are never
limited.

Running in cgroup "foo" without any i/o limitation:
$ echo $$ | sudo tee /cgroup/foo/tasks
$ ./dd-band.sh
===
write:    67108864 bytes (67 MB) copied, 0.917311 s, 73.2 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.437337 s, 153 MB/s
read:     67108864 bytes (67 MB) copied, 2.86287 s, 23.4 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0966508 s, 694 MB/s
===
write:    67108864 bytes (67 MB) copied, 0.893401 s, 75.1 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.447029 s, 150 MB/s
read:     67108864 bytes (67 MB) copied, 2.94607 s, 22.8 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.093962 s, 714 MB/s
===
write:    67108864 bytes (67 MB) copied, 0.896389 s, 74.9 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.437567 s, 153 MB/s
read:     67108864 bytes (67 MB) copied, 2.8686 s, 23.4 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.10307 s, 651 MB/s
===
write:    67108864 bytes (67 MB) copied, 0.896151 s, 74.9 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.436411 s, 154 MB/s
read:     67108864 bytes (67 MB) copied, 3.19625 s, 21.0 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0914293 s, 734 MB/s
===
write:    67108864 bytes (67 MB) copied, 0.896905 s, 74.8 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.436574 s, 154 MB/s
read:     67108864 bytes (67 MB) copied, 2.77595 s, 24.2 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0955861 s, 702 MB/s

The i/o bandwidth of cgroup "foo" is limited up to 8MB/s:
# echo 8192 > /cgroup/foo/blockio.bandwidth
$ ./dd-band.sh
===
write:    67108864 bytes (67 MB) copied, 8.51182 s, 7.9 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.473478 s, 142 MB/s
read:     67108864 bytes (67 MB) copied, 8.19994 s, 8.2 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0886488 s, 757 MB/s
===
write:    67108864 bytes (67 MB) copied, 8.64248 s, 7.8 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.440868 s, 152 MB/s
read:     67108864 bytes (67 MB) copied, 8.19815 s, 8.2 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0956493 s, 702 MB/s
===
write:    67108864 bytes (67 MB) copied, 8.65532 s, 7.8 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.477425 s, 141 MB/s
read:     67108864 bytes (67 MB) copied, 8.20111 s, 8.2 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0978265 s, 686 MB/s
===
write:    67108864 bytes (67 MB) copied, 8.6522 s, 7.8 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.439753 s, 153 MB/s
read:     67108864 bytes (67 MB) copied, 8.21767 s, 8.2 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0991603 s, 677 MB/s
===
write:    67108864 bytes (67 MB) copied, 8.66015 s, 7.7 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.449204 s, 149 MB/s
read:     67108864 bytes (67 MB) copied, 8.27809 s, 8.1 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0874324 s, 768 MB/s

The i/o bandwidth of cgroup "foo" is increased up to 16MB/s:
# echo 16384 > /cgroup/foo/blockio.bandwidth
$ ./dd-band.sh
===
write:    67108864 bytes (67 MB) copied, 4.37976 s, 15.3 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.460999 s, 146 MB/s
read:     67108864 bytes (67 MB) copied, 4.36709 s, 15.4 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0978816 s, 686 MB/s
===
write:    67108864 bytes (67 MB) copied, 4.40452 s, 15.2 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.437004 s, 154 MB/s
read:     67108864 bytes (67 MB) copied, 4.28799 s, 15.7 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0912823 s, 735 MB/s
===
write:    67108864 bytes (67 MB) copied, 4.45284 s, 15.1 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.440571 s, 152 MB/s
read:     67108864 bytes (67 MB) copied, 4.21476 s, 15.9 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0957647 s, 701 MB/s
===
write:    67108864 bytes (67 MB) copied, 4.4364 s, 15.1 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.440202 s, 152 MB/s
read:     67108864 bytes (67 MB) copied, 4.19432 s, 16.0 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0930223 s, 721 MB/s
===
write:    67108864 bytes (67 MB) copied, 4.45086 s, 15.1 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.4487 s, 150 MB/s
read:     67108864 bytes (67 MB) copied, 4.37955 s, 15.3 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0983858 s, 682 MB/s

The i/o bandwidth of cgroup "foo" is reduced to 4MB/s:
# echo 4096 > /cgroup/foo/blockio.bandwidth
$ ./dd-band.sh
==
write:    67108864 bytes (67 MB) copied, 16.783 s, 4.0 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.456829 s, 147 MB/s
read:     67108864 bytes (67 MB) copied, 16.3971 s, 4.1 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.101392 s, 662 MB/s
===
write:    67108864 bytes (67 MB) copied, 16.8167 s, 4.0 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.472562 s, 142 MB/s
read:     67108864 bytes (67 MB) copied, 16.3957 s, 4.1 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0930801 s, 721 MB/s
===
write:    67108864 bytes (67 MB) copied, 16.8834 s, 4.0 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.445081 s, 151 MB/s
read:     67108864 bytes (67 MB) copied, 16.3901 s, 4.1 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0875281 s, 767 MB/s
===
write:    67108864 bytes (67 MB) copied, 16.8157 s, 4.0 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.469043 s, 143 MB/s
read:     67108864 bytes (67 MB) copied, 16.3917 s, 4.1 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.0922022 s, 728 MB/s
===
write:    67108864 bytes (67 MB) copied, 16.9313 s, 4.0 MB/s
re-write: 67108864 bytes (67 MB) copied, 0.510635 s, 131 MB/s
read:     67108864 bytes (67 MB) copied, 16.3983 s, 4.1 MB/s
re-read:  67108864 bytes (67 MB) copied, 0.089941 s, 746 MB/s

Following are reported more complex benchmarks, using iozone to generate
different i/o workloads. In particular all the results below are
evaluated considering read, re-read, write, re-write, random-read and
random-write operations on files ranging from 8MB up to 32MB and record
size ranging from 1K to 8K.

Using more linear i/o workloads (like the simple "dd"s) strongly reduces
errors and deviations respect to the expected performance results, but
the purpose of the test is to actually verify the effectiveness of the
throttling approach in performance predictability, in particular in
presence of heterogeneous and complex workloads.

== Test #1: parallel iozone, balanced load (without cgroup limitations) ==

This test is a run of 2 parallel iozone: 1 in cgroup "foo" and the other
in cgroup "bar", without any bandwidth limitation.

cgroup:foo$ iozone -I -a -i0 -i1 -i2 -i3 -