cinder-specs/specs/victoria/support-volume-local-cache.rst

Support volume local cache

https://blueprints.launchpad.net/cinder/+spec/support-volume-local-cache

This blueprint proposes to add support of volume local cache in cinder and os-brick. Cache software such as open-cas¹ can use fast NVME SSD or persistent memory (configured as block device mode) to cache for slow remote volumes.

Problem description

Currently there are different types of fast NVME SSDs, such as Intel Optane SSD, with latency as low as 10 us. What's more, persistent memory which aim to be SSD size but DRAM speed gets popupar now. Typical latency of persistent memory would be as low as hundreds of nanoseconds. While typical latency of remote volume for a VM can be at the millisecond level (iscsi / rbd). So these fast SSDs or persistent memory can be mounted locally on compute nodes and used as a cache for remote volumes.

In order to do the cache, three projects need to be changed. These are cinder, os-brick, and Nova. Meanwhile related hardware should be added in system, e.g. high performance SSD or persistent memory. The mechanism is similar to volume encryption where dm-crypt is used. Cache software support would be added in os-brick and Nova will call os-brick when it is trying to attach a volume, os-brick then calls the cache software to setup the cache for the volume. After that, a new virtual block device would be created and laying upon the original block device. os-brick will expose this new virtual block device to Nova, meanwhile the original block device mount point will remain unchanged. All of the cache setup and teardown would be handled within os-brick.

A property would be added in the extra specs of the volume type. If a volume type has extra-spec of "cacheable", then it means the related volumes can be cached in compute node locally.

Like all the local cache solution, multi-attach cannot be supported. This is because cache on node1 doesn't know the changes made to backend volume by node2. So cinder should guarantee these two properties cannot be set at the same time.

Considering VM migration, cache software should not format the source volume (the volume to be cached). So cache software such as bcache cannot be used. This spec only supports open-cas. open-cas is easy to use, you just need to specify a block device as the cache device, and then can use this device to cache other block devices. This is transparent to upper layer and lower layer. Regarding upper layer, guest doesn't know it is using an emulated block device. Regarding lower layer, backend volume doesn't know it is cached, and the data in backend volume will not have extra change because of cache. That means even if the cache is lost for some reason, the backend volume can be mounted to other places and become available immediately. open-cas supports the cache modes below:

• Write-Through (WT): write to cache and backend storage at the same time. So the data is fully synced in cache and backend storage. There will not be any data corruption when cache becomes invalid.
• Write-Around (WA): like Write-Through, but only cache for volume blocks that are already in cache.
• Write-Invalidate (WI): write to backend storage and invalidate cache
• Write-Back (WB): write to cache and lazy write to backend storage. This mode has better write performance but is possible to lose data because the latest data is in cache, but may not be flushed to the backend storage.
• Write-Only (WO): like write-back, but only cache write, will not cache read. So this mode also has the possibility to lose data.

The first three modes are suitable for scenarios that data integrity must be guaranteed, every write io is both written to cache and backend storage. In these modes it is just like read-only cache feature in ceph. No operation would be blocked. e.g. VM live migration, snapshot taking, volume backup and others can be done as usual. Cache software can also be changed to others rather than open-cas at any time because there's no dirty data in cache at any time. This is suitable for read intensive scenario, but it will be no benefit for write io because every write io need to go to backend storage everytime for data integrity.

The last two modes are suitable for scenarios that have high requirements for both read/write performance, but low requirements for data integrity. E.g. some test environment. In these two modes, all operations that depend on the backend volume containing full data would not work safely. So VM live migration, snapshot, volume backups, consistency groups, etc, cannot works safely because there may be still dirty data in cache. At least, operator need to stop the disk io by some way and flush the cache(via casadm) before doing these operations. Cache software also cannot be changed to others except all dirty data has been flushed to backend.

Cache mode can be got from cache instance ID via cache admin tool. Nova passes available cache instances ID list to os-brick, so os-brick knows the cache mode of the cache instances. This spec would make os-brick refuse to attach volume with last two cache modes. But cache mode is set outside of OpenStack, Cinder / os-brick cannot control cache mode being modified after volume attached. But it would be well documented that Write-Back (WB) and Write-Only (WO) mode is dangerous and operators should fully know what they are doing when trying to set to these cache modes.

Some storage backends may support discard/TRIM functionality, but cache software doesn't have sense of this. Cache software evicts data based on policy like 'lru'(this is the default policy of open-cas).

Some storage client, compute node, may enable multipath for volumes, e.g. iscsi/fc volumes. Volume local cache would not work with multipath, with the same reason of multi-attach. os-brick detects multipath via function get_volume_paths() and would not set cache for volume with multipath.

No restrictions would be introduced on retyping a 'cacheable' volume. This is because retyping includes two steps: 1) attach new volume 2) detach old volume. So old volume would be released from caching, new volume would be cached.

This is the shared cache, and the number of volumes to be cached is unlimited. A volume be cached does not mean it will occupy space in cache. The volumes with hot IO will consume more cache space, meanwhile volumes with no IO will not occupy any space in cache devce. The data in cache would be evicted when the data getting cold or other volume's IO getting hot.

Use Cases

• In read intensive scenarios, e.g. AI training, VM volume local cache will significantly boost the storage performance (throughput, and especially disk io latency)

Proposed change

In order to do volume local cache:

• In Nova, end user selects a flavor that is advertised as having a volume-local-cache so guest can be landed in a server with cache capability. The end user should expect different performance based on what server flavor was chosen. More information such as error handling, user message, etc, are out of scope of this spec and would be defined in Nova spec².
• In Cinder, volume type that is 'cacheable' should be selected.

It is Cinder which determines and sets the 'cacheable' property. A volume marked as "cacheable" doesn't mean it must be cached, it just means it is eligible to be cached. Nova calls os-brick to set cache for the volume only when the volume has the property of 'cacheable'.

Cache mode is bound to cache instance. So different cache instances can have different cache modes. All cached volumes share same cache mode if they are cached by the same cache instance. The operator can change cache mode dynamically, using cache software management tool. So cache mode setting is out of OpenStack and is not controlled by OpenStack. Operator should not change cache to unsafe mode. os-brick just accepts the cache name and cache instance IDs from Nova.

cache_name identifies which cache software to use, currently it only supports 'opencas'. Nova knows what cache name is, based on which cache software is enabled in compute node.

Each compute node can have more than one cache instances. os-brick can weight each cache instance passed in, by e.g. total cache size, how many free space, etc, via cache admin tool(casadm), and select the best one.

Some storage types support "extend volume" which triggered from cinder side. e.g. via command "cinder extend ...". It works normally when the volume is not "in-use". But if the volume is attached and "in-use", os-brick would not support to extend and just raise NotImplementedError for the volume with cacheable volume_type. This is because open-cas don't support volume extending dynamically in current release. But "Resize Instance" feature which triggered from Nova still can work, because volume will be detached and then re-attached during ""Resize Instance".

The final solution would be like:

Compute Node

+---------------------------------------------------------+
|                                                         |
|                        +-----+    +-----+    +-----+    |
|                        | VM1 |    | VM2 |    | VMn |    |
|                        +--+--+    +--+--+    +-----+    |
|                           |          |                  |
+---------------------------------------------------------+
|                           |          |                  |
| +---------+         +-----+----------+-------------+    |
| |  Nova   |         |          QEMU Virtio         |    |
| +-+-------+         +-----+----------+----------+--+    |
|   |                       |          |          |       |
|   | attach/detach         |          |          |       |
|   |                 +-----+----------+------+   |       |
| +-+-------+         | /dev/cas1  /dev/cas2  |   |       |
| | osbrick +---------+                       |   |       |
| +---------+ casadm  |        open cas       |   |       |
|                     +-+---+----------+------+   |       |
|                       |   |          |          |       |
|                       |   |          |          |       |         Storage
|              +--------+   |          |    +-----+----+  | rbd   +---------+
|              |            |          |    | /dev/sdd +----------+  Vol1   |
|              |            |          |    +----------+  |       +---------+
|        +-----+-----+      |          |                  |       |  Vol2   |
|        | Fast SSD  |      |    +-----+----+   iscsi/fc/...      +---------+
|        +-----------+      |    | /dev/sdc +-------------+-------+  Vol3   |
|                           |    +----------+             |       +---------+
|                           |                             |       |  Vol4   |
|                     +-----+----+    iscsi/fc/...        |       +---------+
|                     | /dev/sdb +--------------------------------+  Vol5   |
|                     +----------+                        |       +---------+
|                                                         |       |  .....  |
+---------------------------------------------------------+       +---------+

Changes would include:

• Add "cacheable" property in extra-spec of volume type

  • Volume local cache cannot work with multiattach. So when adding "cacheable" to extra-spec, cinder should check if "multiattach" property exists or not. If "multiattach" exists, then cinder should refuse to add "cacheable" property to volume type, and vice versa.
  • Fill "cacheable" property in connection_info. So os-brick can know whether a volume can be cached or not.

• Add a common framework for different cache software in os-brick. This framework should be flexible to support different cache software.

  1) A base class - CacheManager would be added and the main functions would be:

  • __init__()

    This function would accept the parameters from Nova. Parameters include:

    root_helper - used for cache software management tools.

    connection_info - containing device path

    cache_name - specify the cache software name, currently only support 'opencas'

    instance_ids - specify cache instances that can be used. os-brick chooses a best one among these instances

  • attach_volume()

    This function would be called by Nova (in function _connect_volume) to setup cache for a volume when it is trying to attach the volume.

  • detach_volume()

    This function would be called by Nova (in function _disconnect_volume) to release the cache when it is trying to detach a volume.

  2) In __init__.py, a map of cache software and its python class would be added. So os-brick can find the correct class based on cache name.

  CACHE_NAME_TO_CLASS_MAP = {

  "opencas": 'os_brick.caches.opencas.OpenCASEngine', ...

  }

  Meanwhile a function like _get_engine() would be added to go through the map to find the correct class.

• Add the support for open-cas in os-brick.

  Implement functions attach_volume/detach_volume for open-cas.

Code work flow would be like:

Nova                                        osbrick

                                  +
+                                    |
|                                    |
v                                    |
attach_volume                              |
+                                    |
|                                    |
+                                    |
attach_cache                           |
 +                                |
 |                                |
 +                                |
+-------+ volume_with_cache_property?        |
|               +                            |
| No            | Yes                        |
|               +                            |
|     +--+Host_with_cache_capability?        |
|     |         +                            |
|     | No      | Yes                        |
|     |         |                            |
|     |         +-----------------------------> attach_volume
|     |                                      |        +
|     |                                      |        |
|     |                                      |        +
|     |                                      |      set_cache_via_casadm
|     |                                      |        +
|     |                                      |        |
|     |                                      |        +
|     |                                      |      return emulated_dev_path
|     |                                      |        +
|     |                                      |        |
|     |         +-------------------------------------+
|     |         |                            |
|     |         v                            |
|     |   replace_device_path                |
|     |         +                            |
|     |         |                            |
v     v         v                            |
                                  |
attach_encryptor and                          |
rest of attach_volume                         +

• Volume local cache lays upon encryptor would have better performance, but expose decrypted data in cache device. So based on security consideration, cache should lay under encryptor in Nova implementation.

Alternatives

• Assign local SSD to a specific VM. VM can then use bcache internally against the ephemeral disk to cache their volume if they want.

  The drawbacks may include:

  • Can only accelerate one VM. The fast SSD capability cannot be shared by other VMs. Unlike RAM, SSD normally is in TB level and large enough to cache for all the VMs in one node.
  • The owner of the VM should setup cache explicitly. But not all the VM owners want to do this, and not all the VM owners have the knowledge to do this. But they for sure want that the volume performance to be better by default.

• Create a dedicated cache cluster. Mount all the cache (NVME SSD) in the cache cluster as a big cache pool. Then allocate a certain amount of cache to a specific volume. The allocated cache can be mounted on compute node through NVMEoF protocol. Then use cache software to do the same cache.

  But this would be the compete between local PCIe and remote network. The disadvantage of doing it like this is: the network of the storage server would be a bottleneck.

  • Latency: Storage cluster typically provides volumes through iscsi/fc protocol, or through librbd if ceph is used. The latency would be at the millisecond level. Even with NVME over TCP, the latency would be hundreds of microseconds, depending on the network topology. In contrast, the latency of NVME SSD would be around 10 us, take Intel Optane SSD p4800x as example.

• Cache can be added in backend storage side, e.g. in ceph. Storage server normally has its own cache mechanism, e.g. using memory as cache, or using NVME SSD as cache.

  Similiar with above solution, latency is the disadvantage.

REST API impact

None

Data model impact

None

Security impact

• Cache software will remove the cached volume data from cache device when volume is detached. But normally it would not erase the related sectors in cache device. So in theory the volume data is still in cache device before it is overwritten. Unless the cache device is plugged out, otherwise it is acceptable because the volume itself is also mounted and visible on host OS. Volume with encryption doesn't have this issue if encryption laying upon volume local cache.

Notifications impact

None

Other end user impact

None

Performance Impact

• Latency of VM volume would be reduced

Other deployer impact

• Need to configure cache software in compute node

Developer impact

• The support for other cache software can be added by other developers later

Implementation

Assignee(s)

Primary assignee:

Liang Fang <liang.a.fang@intel.com>

Work Items

• Implement a common framework for supporting different cache software
• Support open-cas
• Unit test be added

Dependencies

None

Testing

• Unit-tests, tempest and other related tests will be implemented.
• Test case in particular: leverage DRAM to simulate fast ssd, act as the cache for open-cas; Use fio to do the 4k block size rand read test; Compare the result of volume with / without cache. The expected behavior is: cached volume would get lower latency.

Documentation Impact

• Documentation will be needed. User documentation on how to use cache software to cache volume.

References

[1] https://review.opendev.org/#/c/663549/ [2] https://review.opendev.org/#/c/663542/ [3] https://review.opendev.org/#/c/700799/

---

1. https://open-cas.github.io/↩︎

2. https://review.opendev.org/#/c/689070/↩︎