User Tools

Site Tools


pmem_in_qemu

Quick Summary

QEMU provides excellent options for simulating persistent memory configurations. For all the details about the various QEMU command line options, please check out their documentation:

https://github.com/qemu/qemu/blob/master/docs/nvdimm.txt

In my setup I use the following options to define my NVDIMMs:

-object memory-backend-file,id=mem1,share,mem-path=/home/rzwisler/vms/nvdimm-2,size=17G,align=128M
-device nvdimm,memdev=mem1,id=nv1,label-size=2M

Multiple DIMMs can be defined by varying the id and mem-path object options, and the memdev and id device options. i.e.:

-object memory-backend-file,id=mem1,share,mem-path=/home/rzwisler/vms/nvdimm-3.1,size=17G,align=128M 
-device nvdimm,memdev=mem1,id=nv1,label-size=2M

-object memory-backend-file,id=mem2,share,mem-path=/home/rzwisler/vms/nvdimm-3.2,size=17G,align=128M 
-device nvdimm,memdev=mem2,id=nv2,label-size=2M

According to the QEMU documentation the “align” option was introduced in QEMU v2.12.0, so depending on your distro you may need to install an updated version of QEMU from source.

More Details

Of the various options used in the QEMU command lines above, the following are worth noting:

label-size=2M

This is necessary so that the virtual NVDIMMs have label space. Without label space our DIMMs are used in label-less mode, which is a more restricted configuration that prevents us from creating multiple namespaces per region.

We choose a 2 MiB label size because the label area is consumed out of the total size of the DIMM itself (17 GiB in my configuration), and keeping it to a multiple of 2 MiB allows the resulting namespaces to still use 2 MiB pages in filesystem DAX and device DAX configurations.

align=128M

This option controls the starting alignment of the physical addresses consumed by the NVDIMMs. Because we are using 2 MiB of the 17 GiB of the total space of our NVDIMMs for label space, that means that our NVDIMMs are each (17 GiB - 2 MiB) in size. By default QEMU will put the DIMMs directly next to one another in physical address space, which means that the boundary between them won't align to the 128 MiB memory section size imposed by the Linux kernel.

In short, Linux needs the physical addresses for each NVDIMM to start on a 128 MiB boundary. We can do this by not having a label area so the total physical size of our NVDIMM is 1 GiB aligned, we can have our label space be 128 MiB sized, or we can use the align command line option to manually specify an alignment. Giving up the label area means that we can't have more than a single namespace per region, reducing our configuration flexibility. Having a 128 MiB sized label area means we create label space we will never need and in my experience slows down our VM because the label space needs to be initialized.

pmem_in_qemu.txt · Last modified: 2018/06/01 20:56 by Ross Zwisler