nbdkit’s evil filter

If you want to simulate how your filesystem behaves with a bad drive underneath you have a few options like the kernel dm-flakey device, writing a bash nbdkit plugin, kernel fault injection or a few others. We didn’t have that facility in nbdkit however so last week I started the “evil filter”.

The evil filter can add data corruption to an existing nbdkit plugin. Types of corruption include “cosmic rays” (ie. random bit flips), but more realisticly it can simulate stuck bits. Stuck bits are the only failure mode I can remember seeing in real disks and RAM.

One challenge with writing a filter like this is to make the stuck bits persistent across accesses, without requiring us to maintain a large bitmap in the filter keeping track of their location. The solution is fairly elegant: split the underlying disk into blocks. When we read from a block, reconstruct the stuck bits within that block from a fixed seed (calculated from a global PRNG seed + the block’s offset), and iterate across the block incrementing by random intervals. The intervals are derived from the block’s seed so they are the same each time they are calculated. We size the blocks so that each one will have about 100 corrupted bits so this reconstruction doesn’t take very long. Nothing is stored except one global PRNG seed.

The filter isn’t upstream yet but hopefully it can be another way to test filesystems and distributed storage in future.

Removing the cache from an LV

If you’ve cached an LV (see yesterday’s post), how do you remove the cache and go back to just a simple origin LV?

It turns out to be simple, but you must make sure you are removing the cache pool (not the origin LV, not the CacheMetaLV):

# lvremove vg_guests/lv_cache
  Flushing cache for testoriginlv.
  0 blocks must still be flushed.
  Logical volume "lv_cache" successfully removed

This command deletes the CacheDataLV and CacheMetaLV. To reenable the cache you have to go through the process here again.

This is also what you must (currently) do if you want to resize the LV — ie. remove the cache pool, resize the origin LV, then recreate the cache on top. Apparently they’re going to get lvresize to do the right thing in a future version of LVM.

Using LVM’s new cache feature

If you have a machine with slow hard disks and fast SSDs, and you want to use the SSDs to act as fast persistent caches to speed up access to the hard disk, then until recently you had three choices: bcache and dm-cache are both upstream, or Flashcache/EnhanceIO. Flashcache is not upstream. dm-cache required you to first sit down with a calculator to compute block offsets. bcache was the sanest of the three choices.

But recently LVM has added caching support (built on top of dm-cache), so in theory you can take your existing logical volumes and convert them to be cached devices.

The Set-up

To find out how well this works in practice I have added 3 disks to my previously diskless virtualization cluster:

There are two 2 TB WD hard disks in mirrored configuration. Those are connected by the blue (“cold”) wires. And on the left there is one Samsung EVO 250 GB SSD, which is the red (“hot”) drive that will act as the cache.

In other news: wow, SSDs from brand manufacturers are getting really cheap now!

In the lsblk output below, sda and sdb are the WD hard drives, and sdc is the Samsung SSD:

# lsblk
NAME                                     MAJ:MIN RM   SIZE RO TYPE  MOUNTPOINT
sda                                        8:0    0   1.8T  0 disk  
└─sda1                                     8:1    0   1.8T  0 part  
  └─md127                                  9:127  0   1.8T  0 raid1 
sdb                                        8:16   0   1.8T  0 disk  
└─sdb1                                     8:17   0   1.8T  0 part  
  └─md127                                  9:127  0   1.8T  0 raid1 
sdc                                        8:32   0 232.9G  0 disk  
└─sdc1                                     8:33   0 232.9G  0 part  

Performance

Before starting to set up the caching layer, let’s find out how fast the hard disks are. Note that these figures include the ext4 and LVM overhead (ie. they are done on files on a filesystem, not on the raw block devices). I also used O_DIRECT.

HDD writes: 114 MBytes/sec
HDD reads: 138 MBytes/sec
SSD writes: 157 MBytes/sec
SSD reads: 197 MBytes/sec

Note these numbers don’t show the real benefit of SSDs — namely that performance doesn’t collapse as soon as you randomly access the disk.

Terminology

The lvmcache(7) [so new there is no copy online yet] documentation defines various terms that I will use:

origin LV           OriginLV      large slow LV
cache data LV       CacheDataLV   small fast LV for cache pool data
cache metadata LV   CacheMetaLV   small fast LV for cache pool metadata
cache pool LV       CachePoolLV   CacheDataLV + CacheMetaLV
cache LV            CacheLV       OriginLV + CachePoolLV

Creating the LVs

Since the documentation contains a frankly rather scary and confusing section about all the ways that removing the wrong LV will completely nuke your OriginLV, for the purposes of testing I created a dummy OriginLV with some dummy disk images on the slow HDDs:

# lvcreate -L 100G -n testoriginlv vg_guests
  Logical volume "testoriginlv" created
# mkfs -t ext4 /dev/vg_guests/testoriginlv

Also note that resizing cached LVs is not currently supported (coming later — for now you can work around it by removing the cache, resizing, then recreating the cache).

Creating the cache layer

What is not clear from the documentation is that everything must be in a single volume group. That is, you must create a volume group which includes both the slow and fast disks — it simply doesn’t work otherwise.

Therefore my first step is to extend my existing VG to include the fast disk:

# vgextend vg_guests /dev/sdc1
  Volume group "vg_guests" successfully extended

I create two LVs on the fast SSD. One is the CacheDataLV, which is where the caching takes place. The other is the CacheMetaLV which is used to store an index of the data blocks that are cached on the CacheDataLV. The documentation says that the CacheMetaLV should be 1/1000th of the size of the CacheDataLV, but a minimum of 8MB. Since my total available fast space is 232GB, and I want a 1000:1 split, I choose a generous 1GB for CacheMetaLV, 229G for CacheDataLV, and that will leave some left over space (my eventual split turns out to be 229:1).

# lvcreate -L 1G -n lv_cache_meta vg_guests /dev/sdc1
  Logical volume "lv_cache_meta" created
# lvcreate -L 229G -n lv_cache vg_guests /dev/sdc1
  Logical volume "lv_cache" created
# lvs
  LV                     VG        Attr       LSize
  lv_cache               vg_guests -wi-a----- 229.00g
  lv_cache_meta          vg_guests -wi-a-----   1.00g
  testoriginlv           vg_guests -wi-a----- 100.00g
# pvs
  PV         VG        Fmt  Attr PSize   PFree  
  /dev/md127 vg_guests lvm2 a--    1.82t 932.89g
  /dev/sdc1  vg_guests lvm2 a--  232.88g   2.88g

(You’ll notice that my cache is bigger than my test OriginLV, but that’s fine as once I’ve worked out all the gotchas, my real OriginLV will be over 1 TB).

Why did I leave 2.88GB of free space in the PV? I’m not sure actually. However the first time I did this, I didn’t leave any space, and the lvconvert command [below] complained that it needed 256 extents (1GB) of workspace. See Alex’s comment below.

Convert the CacheDataLV and CacheMetaLV into a “cache pool”:

# lvconvert --type cache-pool --poolmetadata vg_guests/lv_cache_meta vg_guests/lv_cache
  Logical volume "lvol0" created
  Converted vg_guests/lv_cache to cache pool.

Now attach the cache pool to the OriginLV to create the final cache object:

# lvconvert --type cache --cachepool vg_guests/lv_cache vg_guests/testoriginlv
  vg_guests/testoriginlv is now cached.

Benchmark

Looks good, but how well does it work? I repeated my benchmarks above on the cached LV:

LV-cache writes: 114 MBytes/sec
LV-cache reads: 138 MBytes/sec

Which is exactly the same as the backing hard disk.

Luckily this is correct behaviour. Mike Snitzer gave me an explanation of why my test using dd isn’t a useful test of dm-cache.

What I’m going to do next is to start setting up guests, and check the performance inside each guest (which is what in the end I care about).