In an environment with heterogenious worker nodes, a universal resource
table for all workers does not allow effective scheduling of tasks. Some
workers may have different proof cache settings, changing the required
memory for different tasks. Some workers may have a different count of
CPUs per core-complex, changing the max parallelism of PC1.
This change allows workers to customize these parameters with
environment variables. A worker could set the environment variable
PC1_MIN_MEMORY for example to customize the minimum memory requirement
for PC1 tasks. If no environment variables are specified, the resource
table on the miner is used, except for PC1 parallelism.
If PC1_MAX_PARALLELISM is not specified, and
FIL_PROOFS_USE_MULTICORE_SDR is set, PC1_MAX_PARALLELSIM will
automatically be set to FIL_PROOFS_MULTICORE_SDR_PRODUCERS + 1.
Before this change workers can only be allocated one GPU task,
regardless of how much of the GPU resources that task uses, or how many
GPUs are in the system.
This makes GPUUtilization a float which can represent that a task needs
a portion, or multiple GPUs. GPUs are accounted for like RAM and CPUs so
that workers with more GPUs can be allocated more tasks.
A known issue is that PC2 cannot use multiple GPUs. And even if the
worker has multiple GPUs and is allocated multiple PC2 tasks, those
tasks will only run on the first GPU.
This could result in unexpected behavior when a worker with multiple
GPUs is assigned multiple PC2 tasks. But this should not suprise any
existing users who upgrade, as any existing users who run workers with
multiple GPUs should already know this and be running a worker per GPU
for PC2. But now those users have the freedom to customize the GPU
utilization of PC2 to be less than one and effectively run multiple PC2
processes in a single worker.
C2 is capable of utilizing multiple GPUs, and now workers can be
customized for C2 accordingly.
Attempting to report "memory used by other processes" in the MemReserved
field fails to take into account the fact that the system's memory used
includes memory used by ongoing tasks.
To properly account for this, worker should report the memory and swap
used, then the scheduler that is aware of the memory requirements for a
task can determine if there is sufficient memory available for a task.
Worker processes may have memory limitations imposed by Systemd. But
/proc/meminfo shows the entire system memory regardless of these limits.
This results in the scheduler believing the worker has the entire system
memory avaliable and the worker being allocated too many tasks.
This change attempts to read cgroup memory limits for the worker
process. It supports cgroups v1 and v2, and compares cgroup limits
against the system memory and returns the most conservative values to
prevent the worker from being allocated too many tasks and potentially
triggering an OOM event.
This commit removes badger from the deal-making processes, and
moves to a new architecture with the dagstore as the cental
component on the miner-side, and CARv2s on the client-side.
Every deal that has been handed off to the sealing subsystem becomes
a shard in the dagstore. Shards are mounted via the LotusMount, which
teaches the dagstore how to load the related piece when serving
retrievals.
When the miner starts the Lotus for the first time with this patch,
we will perform a one-time migration of all active deals into the
dagstore. This is a lightweight process, and it consists simply
of registering the shards in the dagstore.
Shards are backed by the unsealed copy of the piece. This is currently
a CARv1. However, the dagstore keeps CARv2 indices for all pieces, so
when it's time to acquire a shard to serve a retrieval, the unsealed
CARv1 is joined with its index (safeguarded by the dagstore), to form
a read-only blockstore, thus taking the place of the monolithic
badger.
Data transfers have been adjusted to interface directly with CARv2 files.
On inbound transfers (client retrievals, miner storage deals), we stream
the received data into a CARv2 ReadWrite blockstore. On outbound transfers
(client storage deals, miner retrievals), we serve the data off a CARv2
ReadOnly blockstore.
Client-side imports are managed by the refactored *imports.Manager
component (when not using IPFS integration). Just like it before, we use
the go-filestore library to avoid duplicating the data from the original
file in the resulting UnixFS DAG (concretely the leaves). However, the
target of those imports are what we call "ref-CARv2s": CARv2 files placed
under the `$LOTUS_PATH/imports` directory, containing the intermediate
nodes in full, and the leaves as positional references to the original file
on disk.
Client-side retrievals are placed into CARv2 files in the location:
`$LOTUS_PATH/retrievals`.
A new set of `Dagstore*` JSON-RPC operations and `lotus-miner dagstore`
subcommands have been introduced on the miner-side to inspect and manage
the dagstore.
Despite moving to a CARv2-backed system, the IPFS integration has been
respected, and it continues to be possible to make storage deals with data
held in an IPFS node, and to perform retrievals directly into an IPFS node.
NOTE: because the "staging" and "client" Badger blockstores are no longer
used, existing imports on the client will be rendered useless. On startup,
Lotus will enumerate all imports and print WARN statements on the log for
each import that needs to be reimported. These log lines contain these
messages:
- import lacks carv2 path; import will not work; please reimport
- import has missing/broken carv2; please reimport
At the end, we will print a "sanity check completed" message indicating
the count of imports found, and how many were deemed broken.
Co-authored-by: Aarsh Shah <aarshkshah1992@gmail.com>
Co-authored-by: Dirk McCormick <dirkmdev@gmail.com>
Co-authored-by: Raúl Kripalani <raul@protocol.ai>
Co-authored-by: Dirk McCormick <dirkmdev@gmail.com>
