Performance decrease using Hyper-V Win10 clients

[jkozee]

@Tom-Elliott Ah, yes for ntfs. So, perhaps the block driver.

[jkozee]

@Tom-Elliott I kicked off a script to build the kernels. Assuming they build and boot, I’ll report my findings.

[jkozee]

@Tom-Elliott Hmm. 3.3.2 built but wouldn’t boot. I got a kernel panic, not sycning VFS. I used the config from https://svn.code.sf.net/p/freeghost/code/trunk/kernel/TomElliott.config.64. Is there another one I should use for 3.3.2?

[Tom Elliott]

@jkozee 3.3 is very old. I thought 4.3 worked and 4.4 doesn’t so I would suspect somewhere between those would be enough to start to figure out.

[jkozee]

@Tom-Elliott Um, yeah. That’s what I get for trying to multitask and trying to script the builds. Let me see what I’m actually doing. Sorry for wasting space here…

[jkozee]

@Tom-Elliott Ok, I think I scripted builds 4.3.2 to 4.3.5 and 4.4.1, but I’ll start over just to be sure. I see the config for 4.3 on the repo at r4316. Let me start there and see what I get…

[Tom Elliott]

@jkozee look on wiki for build tomelliott kernel

Follow instructions and please test with the additional patches. Speed up build time by adding -j $(nproc) to the make commands

[Sebastian Roth]

@jkozee said:

I see the config for 4.3 on the repo at r4316. Let me start there and see what I get…

Don’t bother too much about getting the exact config Tom used for a particular version. I’d suggest using the newest config for all the builds. As far as I know - hope this is correct - make oldconfig will ask you on the console if there are settings missing. Older ones will just be tossed.

As well, using the same config as a base is wise to properly compare the different kernels versions. Otherwise you end up wondering if a change in config made the difference!

[jkozee]

@Sebastian-Roth
Looks like my script wasn’t copying the .config file, so I was building with the defaults.

I updated it and it built 4.3.2 and it boots fine now. I used the latest config and my script does “yes ‘’ | make oldconfig”. I’ll let it build the ones I mentioned earlier and test them. I’m about out of time for now, so I’ll post the results later.

@Tom-Elliott
I did not have time to write a sed script to include the additional patches from the wiki, but I can do that later or apply them by hand, once I have a chance to test the scripted builds.

Sound reasonable?

[jkozee]

Build script finished quicker than I expected. Looks like it was introduced between 4.3.5 and 4.4.1. I’ll look at git bisect when I can make the time.

[Sebastian Roth]

@jkozee Great work! I am sure you will see what’s exactly causing it and when it was introduced! bisect is your friend.

[jkozee]

@Sebastian-Roth and @Tom-Elliott

The change to the kernel is actually in the scsi driver.

The commit that introduced the delay is 81988a0e6b031bc80da15257201810ddcf989e64, which applies changes to drivers/scsi/storvsc_drv.c.

I can confirm that reverting the diff on 4.4.2 brings the performance on the hyper-v client on par with 4.3.2. I can’t speak to the commit itself, as I just blindly reverted it and didn’t spend any time on digesting the patch itself.

My timings on the patched 4.4.2 was 2:14 for the deploy and 18:20 for the capture. That means the deploy is 50% faster and the capture is 27% slower than my tests for 4.3.2. @Tom-Elliott I did not include the additional patches you mentioned either, so I would need to retest both kernels under the same server conditions (and with the additional patches applied to 4.4.2) for more accurate results.

Hope this proves useful.

[jkozee]

Still seems more like the issue should be with the block device, rather than the scsi driver. Seems like it would be related to caching or block size/block alignment of the ssd.

[jkozee]

To me, these lines from the commit look most interesting:
/* Ensure there are no gaps in presented sgls */
blk_queue_virt_boundary(sdevice->request_queue, PAGE_SIZE - 1);

[jkozee]

So, adding that line to 4.3.2 results in performance degradation and removing it from 4.4.2 results in performance increase. Guess all that’s left is to figure out (understand) what it actually does…

[jkozee]

Maybe this makes more sense: blk_queue_virt_boundary(sdevice->request_queue, sdevice->page_size - 1);

[Sebastian Roth]

@jkozee You are great man! This is what I love about open source and the people knowing how to go with it…

I didn’t even know that there is a storage driver for Hyper-V (and VMware by the way) right in the linux kernel. And we have it enabled: CONFIG_HYPERV_STORAGE

Looking at the scsi_device (sdevice) struct I don’t see page_size. So I don’t think your change is gonna work. Probably wouldn’t compile at all. Looking through the driver code I see PAGE_SIZE used several times. So I feel like this seams ok - although I don’t know much about this particular driver!

You just might want to get in contact with the authors of the storvsc_drv driver (line 18ff) and as well the author of the patch! Tell them that you bisected a major slowdown issue to that particular patch. I guess that their input is a lot more helpful than trying to understand the whole driver by yourself! Please keep us posted.

Edit: Seams like the function blk_queue_virt_boundary was added only a few months back - intended to be used with NVMe devices from what it looks to me (I don’t think this has anything to do with you seeing slower speeds on the SSD backend though).

Edit2: Further interesting things to read on this are here and here - infiniband driver using this as well, interesting part is:

(Very) nice cleanup – so what’s the actual deal here? is that as long
as we plant a slave alloc callback into out scsi host template which
further invokes a
blk_queue_virt_boundary(sdev->request_queue, ~MASK_4K) call, we’re
100% safe/sure what all SGs we get meet the alignment criteria?

Correct, the nvme driver has the same alignment constraints for its PRPs
and uses the queue virt_boundary to have the block layer enforce the
SG alignment.

[jkozee]

@Sebastian-Roth Oops, yes page_size is not part of the sdevice struct. It would probably be more appropriate to rollback the 81988a0e6b031bc80da15257201810ddcf989e64 anyhow. Leaving blk_queue_virt_boundary set to 0, rather than setting it to PAGE_SIZE-1 appears to fix the slowdown, but it would to take some research to determine what other impact that might have. I’ll probably just revert to 4.3.2 for my VMs until I have more time to investigate the issue.

Edit:
In fact, it looks like “Linux Integration Services for Microsoft Hyper-V” also diverge from the Bounce buffer commit: https://github.com/LIS/lis-next/blob/master/hv-rhel6.x/hv/storvsc_drv.c

Looks like at least one of the authors of storvsc_drv.c is on the project, but not active.

[jkozee]

I reached out to the author of the patch. I’ll post if any new information becomes available.

[Tom Elliott]

@jkozee I’ve added a bit of code to the found file that seems to be causing the issue.

If you would like to try it, it can be downloaded at:

http://mastacontrola.com/bzImage (64bit)

I only built the 64 bit kernel.

The adjusted code is:

--- a/linux/drivers/scsi/storvsc_drv.c 2016-02-19 09:46:33.272075454 -0500
+++ b/linux/drivers/scsi/storvsc_drv.c 2016-02-23 17:23:12.868518253 -0500
@@ -1231,7 +1231,8 @@
        blk_queue_rq_timeout(sdevice->request_queue, (storvsc_timeout * HZ));
 
        /* Ensure there are no gaps in presented sgls */
-       blk_queue_virt_boundary(sdevice->request_queue, PAGE_SIZE - 1);
+    if (PAGE_SIZE - 1 < 0) blk_queue_virt_boundary(sdevice->request_queue, 0);
+    else blk_queue_virt_boundary(sdevice->request_queue, PAGE_SIZE - 1);
 
        sdevice->no_write_same = 1;

It’s all on theory that what’s happening is it’s putting the disk to a negative number, and this causes the slowdown. Again, probably won’t work, but still would be nice to know for sure.