[RFC] qemu: set maxmem to host memory size #336

wcwxyz · 2017-08-09T06:36:02Z

Setting maxmem to host memory size wouldn't change memory footprint. It
only defines the upper bounds for memory hotplug.

Currently we set maxmem to mem*1.5, which is not enough for running a
small VM (eg. 128M) considering nvdimm device consumes memory address
space.

This will also allow us to do memory hotplug (maybe in the future).

Signed-off-by: WANG Chao [email protected]

wcwxyz · 2017-08-09T06:41:03Z

CI failed because I didn't vendor github.com/moby/moby/pkg/system.

Can you guys please show me how to vendor a new go project using dep?

Or should I write a new function (look up /proc/meminfo) to get host memory instead import a new pkg?

sboeuf · 2017-08-09T07:04:42Z

Hi @wcwxyz, thanks again for this new patch :)
I have just looked at it, and I would prefer that you avoid to import this moby package since what we need is very simple. Moreover, we are not trying to support Windows here, so you can directly create our own function to read from /proc/meminfo. This way, your problems are solved ;)

coveralls · 2017-08-09T08:39:23Z

Coverage decreased (-0.05%) to 66.449% when pulling 5dbade4 on wcwxyz:qemu-set-maxmem into 59004f2 on containers:master.

wcwxyz · 2017-08-09T08:47:11Z

I've updated the patch. I noticed coverage drop down a bit. That's because some error case is not tested (but should be ok).

coveralls · 2017-08-09T14:07:20Z

Coverage decreased (-0.05%) to 66.449% when pulling e73fdb3 on wcwxyz:qemu-set-maxmem into 59004f2 on containers:master.

grahamwhaley · 2017-08-09T14:27:50Z

qemu.go

+		return ciaoQemu.Memory{}, fmt.Errorf("Unable to read memory info: %s", err)
+	}
+	// add 1G memory space for nvdimm device (vm guest image)
+	memMax := fmt.Sprintf("%dM", int(float64(hostMem)/1024/1024)+1024)


in the spirit of @jodh-intel , it might be nice if the variable names reflected the units they were holding, maybe like memMaxMegs or similar. not a blocker :-)

emm... the code itself is clear enough. Since memMax is a string, I think I'm gonna change hostMem to hostMemBytes to be clearer.

grahamwhaley · 2017-08-09T14:30:23Z

lgtm
let's add a:
Fixes #337

jodh-intel · 2017-08-09T15:07:02Z

qemu.go

+	return memory, nil
+}
+
+func getHostMemorySize() (uint64, error) {


I'd prefer this was called getHostMemorySizeBytes() to avoid any confusion.

jodh-intel · 2017-08-09T15:09:53Z

qemu.go

+}
+
+func getHostMemorySize() (uint64, error) {
+	f, err := os.Open("/proc/meminfo")


This function really need a test so I'd either pass this filename as an argument, or use a global that the tests can modify. For an almost identical example (feel free to copy):

https://github.com/clearcontainers/runtime/blob/master/cc-check.go#L85..L99

https://github.com/clearcontainers/runtime/blob/master/cc-check_test.go#L45..L90

jodh-intel · 2017-08-09T15:10:50Z

qemu.go

+		return 0, err
+	}
+
+	return 0, fmt.Errorf("unable to get MemTotal in /proc/meminfo")


Again, this message can reference the variable for /proc/meminfo.

jodh-intel · 2017-08-09T15:11:18Z

qemu_test.go

@@ -395,10 +395,16 @@ func TestQemuSetMemoryResources(t *testing.T) {

 	q := &qemu{}

+	hostMem, err := getHostMemorySize()


s/hostMem/hostMemBytes/

Sure. Thanks.

jodh-intel · 2017-08-09T15:12:03Z

qemu_test.go

+	if err != nil {
+		t.Fatal(err)
+	}
+	memMax := fmt.Sprintf("%dM", int(float64(hostMem)/1024/1024)+1024)


I'd add a paranoid check to ensure that hostMemBytes > 0 before the division. It would also be clearer potentially to define variables for the division + addition with comments explaining why those values were chosen.

yes, you're right. Will do that.

coveralls · 2017-08-10T08:32:43Z

Coverage increased (+0.04%) to 66.542% when pulling 524f0d3 on wcwxyz:qemu-set-maxmem into 59004f2 on containers:master.

coveralls · 2017-08-10T08:35:57Z

Coverage increased (+0.04%) to 66.542% when pulling ba7761a on wcwxyz:qemu-set-maxmem into 59004f2 on containers:master.

jodh-intel

lgtm, but I'd like @grahamwhaley to review as well.

jodh-intel · 2017-08-10T09:02:30Z

qemu_test.go

+	file := filepath.Join(dir, "meminfo")
+	// file doesn't exist
+	if _, err := getHostMemorySizeBytes(file); err == nil {
+		t.Fatal()


Nit: You could remove the comment above and make this:

t.Fatalf("expected failure as file %q does not exist", file)

yeah, sure.

jodh-intel · 2017-08-10T09:03:25Z

qemu.go

@@ -106,6 +108,11 @@ const (
 	maxDevIDSize = 31
 )

+const (
+	// NVDIMM device needs memory space (1024MB is suggested by @grahamwhaley)


+1 for the comment but please can you add a link to an issue/PR rather than mentioning individuals.

something like this? clearcontainers/runtime#380 (comment)

I mean the full url, github seems automatically format that.

looks correct to me :-) (yes, github nicely hides the full URL where it can :-) )

I agree with @jodh-intel, you should not mention @grahamwhaley directly into the source code. And please refer only the isssue, not the very specific comment.

wcwxyz · 2017-08-10T09:25:59Z

@jodh-intel thanks for review. I'll update the code after @grahamwhaley's review.

wcwxyz · 2017-08-10T10:14:54Z

I take @grahamwhaley is ok with the rest of the patch.
Re-pushed. Please take a look.

coveralls · 2017-08-10T10:19:19Z

Coverage increased (+0.04%) to 66.542% when pulling 23748cc on wcwxyz:qemu-set-maxmem into 59004f2 on containers:master.

grahamwhaley · 2017-08-10T10:35:24Z

thanks @wcwxyz that:
lgtm
@sboeuf can you review, and if happy, merge?

sboeuf · 2017-08-11T06:40:19Z

I'll review tomorrow, please don't merge until I have reviewed this one.

grahamwhaley · 2017-08-11T10:11:42Z

adding DNM label until @sboeuf gets a chance to review etc.

sboeuf · 2017-08-12T15:46:36Z

qemu.go

@@ -106,6 +108,11 @@ const (
 	maxDevIDSize = 31
 )

+const (
+	// NVDIMM device needs memory space (1024MB is suggested by @grahamwhaley)


I agree with @jodh-intel, you should not mention @grahamwhaley directly into the source code. And please refer only the isssue, not the very specific comment.

sboeuf · 2017-08-12T15:47:01Z

qemu.go

+const (
+	// NVDIMM device needs memory space (1024MB is suggested by @grahamwhaley)
+	// See https://github.com/clearcontainers/runtime/issues/380#issuecomment-321216520
+	defaultNvdimmMemory = 1024


I would call this constant maxMemoryOffset because that's basically what it is, and we don't need to specify the reason in the variable name. As long as we have your comment, a generic name sounds better to me.

sboeuf · 2017-08-12T15:47:13Z

qemu.go

@@ -468,13 +476,20 @@ func (q *qemu) setCPUResources(podConfig PodConfig) ciaoQemu.SMP {
 	return smp
 }

-func (q *qemu) setMemoryResources(podConfig PodConfig) ciaoQemu.Memory {
+func (q *qemu) setMemoryResources(podConfig PodConfig) (ciaoQemu.Memory, error) {
+	hostMemBytes, err := getHostMemorySizeBytes("/proc/meminfo")


/proc/meminfo is a very well known path for Linux. Please define this as a const and don't provide it through this call, instead use it directly from the function getHostMemorySizeBytes().

sboeuf · 2017-08-12T15:47:47Z

qemu.go

+	}
+
+	// add 1G memory space for nvdimm device (vm guest image)
+	memMax := fmt.Sprintf("%dM", int(float64(hostMemBytes)/1024/1024)+defaultNvdimmMemory)


This is wrong to do that, isn't it ? I mean you're clearly asking to run a VM with a memory max greater than what the system has to offer ? Or am I reading the code wrong ?
Instead you should check that the system can allow at least the expected memory to be set and this would be a nice error to throw. IMO, we don't want to continue if our host cannot provide the amount of memory requested, because the user could end up in not being able to run the workload he expects.

And I think that you should validate that the expected amount of memory + the offset can fit in the host memory. But please don't consider this case as throwing an error, just log a warning, because the offset might be too much in some cases that could end up functioning very well. Indeed, in case you have the NVDIMM using 200MiB and you ask 1.5GiB on a 2GiB host, you really need only 1.7 and this will work great. But if we check with our offset of 1GiB, we re gonna fail, that's why we don't want to generate an error in that case.

No. maxmem is used to define upper bounds for memory hotplug. You request memory via qemu -m MEMORY.

Say run qemu -m 2048M,slots=2,maxmem=30000M, you got 2048M for on your VM and you can hot plug in 2 rom device (pc-dimm) into VM, the total memory can not exceed 30000M. In case your host memory is less than 30000M, hot plug will fail when your ask for memory when host doesn't have available.

That's exactly what you're saying, in case the host memory is less than 30G, the hotplug will fail. That's why I think we should try to set a maxmem higher than the total RAM available on the host.

Aha, I understand what you're trying to say.

You're wrong about "offset". It is not consuming real host memory. Nvdimm needs "offset" to work on qemu. It takes memory space, not the real memory. (I'm not familiar with nvdimm device, but a quick test suggests that).

For maxmem, this is just the memory upper bounds for qemu. Setting maxmem to host_total_memory+offset is not asking qemu to run a VM with that much memory. Qemu will accept or reject to proceed any memory hotplug requests based on the maxmem. Whether hot plug works or not, depends kernel memory management.

For 'offset', allocating that extra space and the extra slot does not take any host memory at startup. Then later we plug the NVDIMM into that space. I think we are probably safe not calculating that into the 'total host memory' equation.

For 'mem_request > host_memory' - whether the host allows or denies a VM (CC-container) to be created if it requests more memory than the host has either available or system total I believe actually comes down to how the host kernel is configured for over-commit. If the system owner has knowingly configured the system to allow VM overcommit then we should probably not error. I think we could be more informative though with one or both of the following options maybe:

warn if the requested memory is > the current system free memory (the default kernel non-overcommit failure case I believe)

warn if the requested memory is > the system memory total (because you are likely going to run into performance issues if your container uses all that memory)

For more information on kernel overcommit settings etc. see:
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/tree/Documentation/sysctl/vm.txt
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/tree/Documentation/vm/overcommit-accounting

Thanks @grahamwhaley for elaborating on memory overcommit.

I don't think we should do anything about user requested memory. Runtime should not log error/warn solely based on overcommit or free memory. When QEMU fails to allocate memory, it errors out. There's enough information for user to see.

Kernel memory management is tricky. I really don't think we should make assumptions about memory is not enough for VM under any condition. Because we can only know memory actually runs out when there's allocation failure at the moment.

@grahamwhaley @sboeuf What do you think?

I'm happy with that theory - we cannot tell when a VM launch will fail or not, as we don't have the intimate knowledge of the system.
Later if we find that the qemu failure is non-obvious from the logs, then we can maybe work on improving how the error/logs are gathered and pushed up through the runtime.

Thanks. I believe this part of code is ok, if @sboeuf doesn't have any objections.

Ok you convinced me :)

Thanks for review @sboeuf @grahamwhaley. I'll post v2.

sboeuf · 2017-08-12T15:48:07Z

qemu.go

+		if err != nil {
+			continue
+		}
+		bytes := uint64(size) * 1024


Why are we converting to bytes ? I mean we're basically multiplying by 1024 here to divide twice from the caller. I would not bother doing this and instead name the function getHostMemorySizeKb.

I'll change to KB.

coveralls · 2017-08-15T06:57:02Z

Coverage increased (+0.02%) to 66.527% when pulling d77b1cf on wcwxyz:qemu-set-maxmem into 58ebe95 on containers:master.

sboeuf

Thanks for this v2, the code itself looks good to me, but I would like to see those generic functions/variables as part of hypervisor.go so that they could be reused for other hypervisor implementations in the future.
I am looking forward to the v3 ;)

sboeuf · 2017-08-15T07:09:29Z

qemu.go

+	maxMemoryOffset = 1024
+)
+
+const procMemInfo = "/proc/meminfo"


Please move this to hypervisor.go since it can be reused across different hypervisor implementations.

sboeuf · 2017-08-15T07:12:24Z

qemu.go

+	return memory, nil
+}
+
+func getHostMemorySizeKb(memInfoPath string) (uint64, error) {


Please move this to hypervisor.go since it can be reused across different hypervisor implementations.

Setting maxmem to host memory size wouldn't change memory footprint. It only defines the upper bounds for memory hotplug. Currently we set maxmem to mem*1.5, which is not enough for running a small VM (eg. 128M) considering nvdimm device consumes memory address space. This will also allow us to do memory hotplug (maybe in the future). Fixes #337 Signed-off-by: WANG Chao <[email protected]>

wcwxyz · 2017-08-15T07:47:53Z

Here comes v3. Thanks.

coveralls · 2017-08-15T07:50:11Z

Coverage increased (+0.02%) to 66.527% when pulling d0bb122 on wcwxyz:qemu-set-maxmem into 58ebe95 on containers:master.

wcwxyz changed the title ~~qemu: set maxmem to host memory size~~ [RFC] qemu: set maxmem to host memory size Aug 9, 2017

wcwxyz mentioned this pull request Aug 9, 2017

VM memory size calculation needs deciding #337

Closed

grahamwhaley reviewed Aug 9, 2017

View reviewed changes

grahamwhaley requested a review from sboeuf August 9, 2017 14:28

jodh-intel suggested changes Aug 9, 2017

View reviewed changes

jodh-intel approved these changes Aug 10, 2017

View reviewed changes

grahamwhaley approved these changes Aug 10, 2017

View reviewed changes

grahamwhaley added the do-not-merge label Aug 11, 2017

sboeuf suggested changes Aug 12, 2017

View reviewed changes

sboeuf reviewed Aug 15, 2017

View reviewed changes

sboeuf approved these changes Aug 15, 2017

View reviewed changes

sboeuf removed the do-not-merge label Aug 15, 2017

sboeuf merged commit af35aeb into containers:master Aug 15, 2017

		@@ -395,10 +395,16 @@ func TestQemuSetMemoryResources(t *testing.T) {

		q := &qemu{}

		hostMem, err := getHostMemorySize()

[RFC] qemu: set maxmem to host memory size #336

[RFC] qemu: set maxmem to host memory size #336

Conversation

wcwxyz commented Aug 9, 2017

wcwxyz commented Aug 9, 2017 • edited Loading

sboeuf commented Aug 9, 2017

coveralls commented Aug 9, 2017 • edited Loading

wcwxyz commented Aug 9, 2017

coveralls commented Aug 9, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

grahamwhaley commented Aug 9, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

coveralls commented Aug 10, 2017 • edited Loading

coveralls commented Aug 10, 2017 • edited Loading

jodh-intel left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wcwxyz commented Aug 10, 2017

wcwxyz commented Aug 10, 2017

coveralls commented Aug 10, 2017 • edited Loading

grahamwhaley commented Aug 10, 2017

sboeuf commented Aug 11, 2017

grahamwhaley commented Aug 11, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

coveralls commented Aug 15, 2017 • edited Loading

sboeuf left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wcwxyz commented Aug 15, 2017

coveralls commented Aug 15, 2017 • edited Loading

wcwxyz commented Aug 9, 2017 •

edited

Loading

coveralls commented Aug 9, 2017 •

edited

Loading

coveralls commented Aug 9, 2017 •

edited

Loading

coveralls commented Aug 10, 2017 •

edited

Loading

coveralls commented Aug 10, 2017 •

edited

Loading

coveralls commented Aug 10, 2017 •

edited

Loading

coveralls commented Aug 15, 2017 •

edited

Loading

coveralls commented Aug 15, 2017 •

edited

Loading