inflight-queue optimization and remove fifo #197
Conversation
| @@ -274,7 +272,7 @@ print_interval (udxperf_client_t *client, uint64_t bytes, uint64_t start, uint64 | |||
| byte_snprintf(bps_buf, sizeof bps_buf, bytes / time_sec, 'a'); | |||
| bps_buf[19] = '\0'; | |||
|
|
|||
| printf("[%3d] %6.4f-%6.4f sec %s %s/sec", stream->local_id, (start - client->start_time) / 1000.0, (end - client->start_time) / 1000.0, bytes_buf, bps_buf, stream->cwnd); | |||
| printf("[%3d] %6.4f-%6.4f sec %s %s/sec", stream->local_id, (start - client->start_time) / 1000.0, (end - client->start_time) / 1000.0, bytes_buf, bps_buf); | |||
There was a problem hiding this comment.
Fixes warning of extra printf parameter
| @@ -34,7 +35,7 @@ debug_print_cwnd_stats (udx_stream_t *stream) { | |||
| static void | |||
| debug_print_outgoing (udx_stream_t *stream) { | |||
| if (DEBUG) { | |||
| for (uint32_t s = stream->remote_acked; s < stream->seq; s++) { | |||
| for (uint32_t s = stream->remote_acked; s != stream->seq; s++) { | |||
There was a problem hiding this comment.
matters when sequence numbers wrap
|
|
||
| if (pkt->type & UDX_PACKET_FREE_ON_SEND) { |
There was a problem hiding this comment.
this is free is safe to remove. the only packet type put in the unordered queue are packet UDX_PACKET_TYPE_STREAM_SEND, which is not free-on-send.
| assert(pkt->transmits > 0); | ||
|
|
||
| // debug_printf("%lu > %lu=%d\n", stream->rack_time_sent, pkt->time_sent, stream->rack_time_sent > pkt->time_sent); | ||
|
|
||
| if (!rack_sent_after(stream->rack_time_sent, stream->rack_next_seq, pkt->time_sent, pkt->seq + 1)) { | ||
| continue; | ||
| if (pkt->time_sent > stream->rack_time_sent) { |
There was a problem hiding this comment.
this is the optimization from RFC 8985:
As an optimization, an implementation can choose to check only
segments that have been sent before RACK.xmit_ts. This can be more
efficient than scanning the entire SACK scoreboard, especially when
there are many segments in flight. The implementation can use a
separate doubly linked list ordered by Segment.xmit_ts, insert a
segment at the tail of the list when it is (re)transmitted, and
remove a segment from the list when it is delivered or marked as
lost. In Linux TCP, this optimization improved CPU usage by orders
of magnitude during some fast recovery episodes on high-speed WAN
networks.
| @@ -1289,13 +1284,13 @@ udx_rto_timeout (uv_timer_t *timer) { | |||
|
|
|||
| // rack 6.3 | |||
|
|
|||
| for (uint32_t seq = stream->remote_acked; seq < stream->seq; seq++) { | |||
| for (uint32_t seq = stream->remote_acked; seq != stream->seq; seq++) { | |||
There was a problem hiding this comment.
bug if RTO triggers during sequence number wrap
| @@ -1510,15 +1505,10 @@ process_sacks (udx_stream_t *stream, char *buf, size_t buf_len) { | |||
|
|
|||
| static void | |||
| process_data_packet (udx_stream_t *stream, int type, uint32_t seq, char *data, ssize_t data_len) { | |||
| if (seq == stream->ack && type == UDX_HEADER_DATA) { | |||
| if (seq == stream->ack && type & UDX_HEADER_DATA) { | |||
There was a problem hiding this comment.
A packet could have type DATA | END, such a packet would end up in the out-of-order queue, and then processed like normal.
| if (stream->socket != NULL) { | ||
| update_poll(stream->socket); | ||
| } | ||
|
|
There was a problem hiding this comment.
I think this is redundant - write_wanted |= WRITE_WANT_STATE is set later in process_packet if the packet type was DATA or END
jthomas43
force-pushed
the
remove-fifo
branch
from
3c51d39 to
609f46b
Compare
This PR updates the linked-queue PR