Stream-Rail distribution with addr + port #1411
Conversation
|
@tom95858 Shall I also include the stream name into the calculation to increase the distribution? |
|
I'll add |
|
@narategithub what's the status of this? |
When the stream forward the stream data, it used the peer address to determine the endpoint in the rail to forward the data. This patch adds `port` into the calculation as well to increase the stream data distribution among the endpoints in the rail.
|
@tom95858 the code is done but the |
narategithub
force-pushed
the
stream-rail-dist
branch
from
53aa7a5
to
f147f17
Compare
|
@tom This is ready for your review. |
| @@ -478,6 +491,7 @@ __stream_deliver(struct ldms_addr *src, uint64_t msg_gn, | |||
| .name_len = name_len, | |||
| .data_len = data_len, | |||
| .name = stream_name, | |||
| .name_hash = hash, | |||
There was a problem hiding this comment.
If you're going to put the hash in the message, you could put the entire thing in there and avoid the computation on every message. The port and addr aren't changing. This would also mean that the same hash is used all the way up the chain from sender to consumer.
There was a problem hiding this comment.
The address got changed in the following case.
There is a logic to exclude the use of local address (e.g. 127.0.0.1) as src address here:
This is to prevent the case that the application running on the same host connects to sampler ldmsd using localhost address and make all src in the messages from all nodes being 127.0.0.1. Basically, if the src address is local, do not resolve it and let the next guy resolve it instead.
There was a problem hiding this comment.
So there must be code somewhere else that checks the address and family and if 0 sets it to the address and port of the forwarder? Where is that logic?
There was a problem hiding this comment.
@narategithub In any event, all we're doing here is randomizing the seed to what we hope will be a random hash. So as-usual, we've found ourselves in a hole so we keep digging. Let's just back up and decide how we're going to generate a random hash.
Adding the (hash of) stream name into the endpoint index calculation to increase entropy (and hopefully more even distribution among endpoints).
narategithub
force-pushed
the
stream-rail-dist
branch
from
f147f17
to
45251ba
Compare
|
Closing this PR. This will be a part of adaptive flow control PR. |
When the stream forward the stream data, it used the peer address to determine the endpoint in the rail to forward the data. This patch adds
portinto the calculation as well to increase the stream data distribution among the endpoints in the rail.