Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Highlighted
Beginner
87 Views

Strange L1 and L2 bandwidth with Advisor 2017 Update 1 (roofline)

Hi,

I have installed Advisor 2017 update 1 and I get very numbers for L1 and L2 bandwidth. L2 gives 1.5e+5 GB/s and L1 is so huge I can't see it on the plot. I think that this strange numbers were not there in the beta release of the roofline model. This behavior is on both my laptop and my desktop.

Could you please confirm this is a bug? Is there any workaround?

Best regards,

Francois

0 Kudos
8 Replies
Highlighted
Employee
87 Views

Hello,

What is your CPU model and OS? And, is it possible to share your application to reproduce the issue?

Regards,
Kirill

0 Kudos
Highlighted
Beginner
87 Views

Hi,

Thanks for your help.

I have seen this problem on many computers. I think this problem started to appear on the last update of Advisor that came in late October. It was not there before.

- Dual-Xeon E5-2660v4, Ubuntu 16.04

Dual-Xeon E5-2660v4, Windows 10

- Core i7 Haswell (MacBook Pro), Ubuntu 16.04 on top of Virtual Machine

- Core i7 Haswell (MacBook Pro), Windows 10 on top of Virtual Machine

What order of magnitude should you expect from L1 and L2 bandwidth?

Francois

0 Kudos
Highlighted
Employee
87 Views

Thanks Francois,

We've seen such problems on Broadwell CPUs, this is a bug not yet fixed. On Haswell there shouldn't be a problem, however virtual machine can be problematic. Do you have an ability to run on another machine?

0 Kudos
Highlighted
Beginner
87 Views

Kirill R. (Intel) wrote:

Thanks Francois,

We've seen such problems on Broadwell CPUs, this is a bug not yet fixed. On Haswell there shouldn't be a problem, however virtual machine can be problematic. Do you have an ability to run on another machine?

Hi Kirill,

I should get a E5-1660v3 (Haswell) soon. But a friend of mine had the same problem with a Sandy Bridge (or an Haswell) running Ubuntu Linux.

0 Kudos
Highlighted
Beginner
87 Views

Kirill R. (Intel) wrote:

Thanks Francois,

We've seen such problems on Broadwell CPUs, this is a bug not yet fixed. On Haswell there shouldn't be a problem, however virtual machine can be problematic. Do you have an ability to run on another machine?

Same problem on a laptop with an Haswell i7-4600U with Ubuntu 14.04.5 LTS, kernel 4.4.0-22 and Parallel Studio XE 2017 Update 1.

0 Kudos
Highlighted
Employee
87 Views

We're investigating the problem. Though we can't reproduce it on any platform as you do. What exact steps do you perform?

0 Kudos
Highlighted
Beginner
87 Views

On my Dual-Xeon E5-2660v4 workstation, running CentOS 7 and Parallel Studio XE 2017 Update 1.

- Compile any code, for instance the following code with "icpc -g =std=c++11 -O3 -xHost -qopenmp advisor.cpp -o advisor"

#include <chrono>
#include <cstdio>

int main() {
  int n = 1000000000;

  double* a = new double;
  double* b = new double;
  double* c = new double;
#pragma omp parallel for
  for (int k = 0; k < n; ++k) {
    a = 0.0;
    b = 0.0;
    c = 0.0;
  }

  auto begin = std::chrono::high_resolution_clock::now();
#pragma omp parallel for
  for (int k = 0; k < n; ++k) {
    a = b + c;
  }
  auto end = std::chrono::high_resolution_clock::now();
  double time =
      1.0e-9 *
      std::chrono::duration_cast<std::chrono::nanoseconds>(end - begin).count();
  std::printf("With first touch policy: %7.2f\n", time);

  delete[] c;
  delete[] b;
  delete[] a;


  return 0;
}

- Launch Advisor, create a new project, select the binary and select "Collect information about flops" in the survey trip count part

- Click "Collect" on "Survey Target"

- Select one of the loop (usually the second loop of my program), and then click "Collect" on "Find trip count and flops".

I get:

- DRAM bandwidth: 123.9 GB/s

- L3 bandwidth: 712.62 GB/s

- L2 bandwidth: 1.4e+6 GB/s

- L1 bandwidth: Can't read

0 Kudos
Highlighted
Employee
87 Views

Thank you for all the details. We've found cause of the problem and have preliminary fix.

0 Kudos