Fix unbounded buffering when one node has slower network #792

Merged
lx merged 2 commits from fix-buffering into main 2024-03-28 12:40:28 +00:00
Owner
  • Add mechanism to limit the number of blocks buffered in RAM as they await being sent to storage nodes. Backpressure is applied to the clients when the limit is reached. The size of this buffer can be controled using the config parameter block_ram_buffer_max

  • Add metric block_ram_buffer_free_kb to visualize memory still available for buffers (when it drops to zero, backpressure is applied)

  • Change the network scheduling algorithm to send requests one after the other, instead of least-attained-service that was equivalent to a round-robin and would cause all blocks to be buffered simultaneously

Related: #788

- [x] Add mechanism to limit the number of blocks buffered in RAM as they await being sent to storage nodes. Backpressure is applied to the clients when the limit is reached. The size of this buffer can be controled using the config parameter `block_ram_buffer_max` - [x] Add metric `block_ram_buffer_free_kb` to visualize memory still available for buffers (when it drops to zero, backpressure is applied) - [x] Change the network scheduling algorithm to send requests one after the other, instead of least-attained-service that was equivalent to a round-robin and would cause all blocks to be buffered simultaneously Related: #788
lx added 2 commits 2024-03-27 15:07:43 +00:00
ci/woodpecker/push/debug Pipeline failed Details
ci/woodpecker/pr/debug Pipeline was successful Details
9613f9495e
[fix-buffering] change request sending strategy and fix priorities
remove LAS, priorize new requests but otherwise just do standard queuing
lx changed title from Fix buffering issues (fix #788) to Fix unbounded buffering when one node has slower network (fix #788) 2024-03-27 15:09:19 +00:00
lx force-pushed fix-buffering from 9613f9495e to 85f580cbde 2024-03-27 15:23:01 +00:00 Compare
Author
Owner

Confirmed that this avoids unbounded memory usage growth inducing OOM kill on a test deployment when sending a single 8GB file to an S3 API server on localhost, which had to send to two other storage nodes, one a bit slower than the other.

With the default value of block_ram_buffer_free_kb (256MB), the RAM usage of Garage as measured by RssAnon is stable under 300MB during the transfer, and the overall speed of the transfer is correctly capped to the speed of the slowest internet link between nodes. Without the patch the API server received data faster than it could send back to storage nodes and RAM usage grew to 1GB, 2GB, 3GB, ... and provoked an OOM kill of the daemon.

Confirmed that this avoids unbounded memory usage growth inducing OOM kill on a test deployment when sending a single 8GB file to an S3 API server on localhost, which had to send to two other storage nodes, one a bit slower than the other. With the default value of `block_ram_buffer_free_kb` (256MB), the RAM usage of Garage as measured by RssAnon is stable under 300MB during the transfer, and the overall speed of the transfer is correctly capped to the speed of the slowest internet link between nodes. Without the patch the API server received data faster than it could send back to storage nodes and RAM usage grew to 1GB, 2GB, 3GB, ... and provoked an OOM kill of the daemon.
lx changed title from Fix unbounded buffering when one node has slower network (fix #788) to Fix unbounded buffering when one node has slower network 2024-03-28 12:05:05 +00:00
lx changed title from Fix unbounded buffering when one node has slower network to Fix unbounded buffering when one node has slower network (fix #788) 2024-03-28 12:37:30 +00:00
lx changed title from Fix unbounded buffering when one node has slower network (fix #788) to Fix unbounded buffering when one node has slower network 2024-03-28 12:39:41 +00:00
lx merged commit ecf641d88c into main 2024-03-28 12:40:28 +00:00
lx deleted branch fix-buffering 2024-03-28 12:40:28 +00:00
Sign in to join this conversation.
No description provided.