RIFS-ES Linux Kernel Scheduler Released

[phoronix]

Phoronix: RIFS-ES Linux Kernel Scheduler Released

RIFS-ES V1 has been published as a new interactivity favor scheduler for the Linux kernel...

http://www.phoronix.com/vr.php?view=MTEyMjg

[3766691]

I am the second one.

If a task sleep and wake up very often it can still gain high priority but the priority will be lower than some task like:
int main()
{
int i;
for(i = 0;i < 60;i++)
sleep(1);
}

These(ES feature) ideas are got from System V scheduler. I improve them and bind them with RIFS.
They are actually a very good solution for desktop. I improve them to make sure that these idea can make a better responaive for the task which needs as small latency as they can.

[jwilliams]

For interactivity, it seems to me that the maximum latencies are much more important than the average. And there does not seem to be a clear improvement over CFS maximum latencies in the benchmarks in the email.

[hechacker1]

Looks promising

The average latency seems way better than CFS, and it even beats BFS at high client counts.

What I'm curious is if the throughput gets killed.

On the other hand, BFS seems to do just as well for low client counts, and has already been debugged far more than this.

Good luck to the developer!

Now if we could only finally resurrect the plugable scheduler system for the kernel.

[uid313]

Graphs

This article is useless without graphs.

Benchmark CFS, BFS, RIFS-ES and do graph.

[3766691]

The average latency seems way better than CFS, and it even beats BFS at high client counts.

What I'm curious is if the throughput gets killed.

On the other hand, BFS seems to do just as well for low client counts, and has already been debugged far more than this.

Good luck to the developer!

Now if we could only finally resurrect the plugable scheduler system for the kernel.

plugable scheduler is a comfortable thing
BFS uses real-time scheduling algorithm so with low client counts it can give O(1) lag bound but with higher amount of client count many task will miss the deadline and it is not O(1) lag anymore.

[Vax456]

This article is useless without graphs.

Benchmark CFS, BFS, RIFS-ES and do graph.

A graph comparing BFS and RIFS-ES will probably be helpful since they're about latency for desktops, but comparing CFS isn't going to help because it's all about throughput.

Plus, comparing BFS and RIFS-ES is going to be tough to quantify. How can you make a chart quantifying the desktop experience with BFS against RIFS-ES? Someone's probably going to have to buy an eight-year old desktop and just eyeball the performance and tell the internet what s/he thinks.

[ulenrich]

funny statistic pic

going up from
60 to 200 clients
you will have latency going from
50000 down to 5000

This is funny!
Is this a quantum machine running faster on high loads?

[3766691]

going up from
60 to 200 clients
you will have latency going from
50000 down to 5000

This is funny!
Is this a quantum machine running faster on high loads?

This is the new version
https://rifs-scheduler.googlecode.co...ec-kernel3.4.x

and the second benchmark
https://rifs-scheduler.googlecode.co...Result-002.PNG

Text version
#MONTH *CFS * * RIFS-ES(new) BFS
1 * * * 21 * * *8 * * * * * * * 7
2 * * * 19 * * *8 * * * * * * * 7
4 * * * 4678 * *1402 * * * * * *2625
8 * * * 9111 * *4361 * * * * * *5271
16 * * *17944 * 7589 * * * * * *12611
32 * * *35628 * 21864 * * * * * 25381
64 * * *69205 * 43145 * * * * * 51653
128 * * 93522 * 35553 * * * * * 97354
255 * * 134560 *6288 * * * * * *200071

Edit 2
The new version has fixed one build issue on NON-SMP processors
And sleep-tracking is disabled.

Edit 3
I have to mention one more time.
The algorithm used in BFS is EEVDF Variation
EEVDF is designed for real-time environment
The good point I have mention already is that wich low workload EEVDF can provide O(1) lag. However with bigger load the lag bound is not O(1) anymore.

[ulenrich]

Chen your graphic Result-002.PNG
looks logical regarding the BFS and CFS curves:
BFS going higher at some point, but better performance at first.

But your RIFS curve is just funny:
The higher the load the better the performance.
Or is the RIFS curve a kind of integration level two: a kind of difference to CFS