Add dm-cache documentation for XOSTOR #295
Conversation
Wescoeur
commented
Nov 6, 2024
Wescoeur
commented
| linstor vg set-property xcp-sr-linstor_group_thin_device 0 Cache/OpMode <MODE> | ||
| ``` | ||
|
|
||
| By default `writethrough` mode is used. This mode is only useful to improve read performance. |
There was a problem hiding this comment.
I'm not sure to understand the cache system: are reads already "cached" by design in LINSTOR? ie doesn't wait to read to another node but having the local disk speed for it. Last time I did a bench, this is what I had. So I am probably ignoring something obvious in there.
There was a problem hiding this comment.
DRBD can use the Linux system's buffers. But by design you don't have a cache: the read requests can be sent to the local host currently used and/or on another machine to accelerate parallel requests.
The interest of dm-cache here is to use HDDs (or slow SSDs) on the DRBD level and a faster device like an NVME to cache the most read blocks.
Signed-off-by: Ronan Abhamon <[email protected]>
Wescoeur
force-pushed
the
xostor-dm-cache
branch
from
c3fc599 to
3728ee4
Compare
| @@ -852,3 +852,66 @@ Note: iptables config must also be modified to remove LINSTOR port rules (edit ` | |||
| For more info: | |||
| - https://linbit.com/blog/linstors-auto-evict/ | |||
| - https://linbit.com/drbd-user-guide/linstor-guide-1_0-en/#s-linstor-auto-evict | |||
|
|
|||
| ### How to-enable dm-cache? | |||
There was a problem hiding this comment.
How to enable dm-cache? (typo)
Should we also add a bit of context? What does the dm-cache do?
| This feature is currently experimental and not covered by [Vates Pro Support](https://vates.tech/pricing-and-support). | ||
| ::: | ||
|
|
||
| On each host, create a new PV and VG using your cache devices: |
There was a problem hiding this comment.
On each host, create a new PV and VG using the cache devices:
| vgcreate linstor_group_cache <CACHE_DEVICES> | ||
| ``` | ||
|
|
||
| Then you can enable the cache with few commands using the linstor controller. |
There was a problem hiding this comment.
Then you can enable the cache with a few commands using the linstor controller:
|
|
||
| Then you can enable the cache with few commands using the linstor controller. | ||
|
|
||
| First, verify the group to modify, this last one must start with "xcp-sr-": |
There was a problem hiding this comment.
- Verify the group to modify. It must start with "xcp-sr-":
| linstor storage-pool list | ||
| ``` | ||
|
|
||
| Make sure the primary resource group is configured with cache support and enable the cache on the volume group: |
There was a problem hiding this comment.
- Make sure the primary resource group is configured with cache support and enable the cache on the volume group:
| linstor rd set-property <VOLUME_NAME> Cache/Cachesize <PERCENTAGE> | ||
| ``` | ||
|
|
||
| It's totally arbitrary. You can go up to 20-30% for for VMS with a high write rate. This should be enough to support a significant number of requests. 10% for solicited VMs. Between 1-5% for VMs with a few requests. You can use 100% if you want, for example for a database on a small VDI with a lot of queries. |
There was a problem hiding this comment.
It's arbitrary and you can go up to 20-30% for VMS with a high write rate.
| It's totally arbitrary. You can go up to 20-30% for for VMS with a high write rate. This should be enough to support a significant number of requests. 10% for solicited VMs. Between 1-5% for VMs with a few requests. You can use 100% if you want, for example for a database on a small VDI with a lot of queries. | ||
|
|
||
| :::warning | ||
| The use of snapshots can consume more memory than necessary due to VHD chains that are too long. It's advisable to limit their use except via XOA during backup processes. |
There was a problem hiding this comment.
Due to too-long VHD chains, snapshots can consume more memory than necessary. It's advisable to limit their use to backup processes via XOA.
| The use of snapshots can consume more memory than necessary due to VHD chains that are too long. It's advisable to limit their use except via XOA during backup processes. | ||
| ::: | ||
|
|
||
| #### How to switch between read and read-write mode? |
There was a problem hiding this comment.
How to switch between read and read-write modes? (typo)
| linstor vg set-property xcp-sr-linstor_group_thin_device 0 Cache/OpMode <MODE> | ||
| ``` | ||
|
|
||
| By default `writethrough` mode is used. This mode is only useful to improve read performance. |
There was a problem hiding this comment.
This mode is only useful for improving read performance.
| This data block is moved later, and the process caller (here tapdisk) is only notified when the block is flushed in the cache disk. | ||
|
|
||
| This algorithm is efficient in not having to wait for writes to be flushed to the local disk as well as to other DRBDs replicated on other nodes. | ||
| But if you have a power outage on the machine using the cache and if it contains data, then the data is lost. |
There was a problem hiding this comment.
However, if a power outage occurs on a machine using a cache that contains data, the data will be lost.
