networks [851f079b]
Tags: Computers
2. Outages
2. Outages
3. Tofu networks for supercomputing
3. Tofu networks for supercomputing
- ieeexplore.ieee.org/document/6041538
- 16d sync for computers
4. linux Monitoring with Pcap
4. linux Monitoring with Pcap
5. Cache Line optimization for cpu
5. Cache Line optimization for cpu
- lore.kernel.org/netdev/20231129072756.3684495-1-lixiaoyan@google.com/
- Up to 45% CPU performance increase by aligning cache lines
6. TCP Congestion Control
6. TCP Congestion Control
- slow start - begins transmission with a small congestion window and exponetnially increases it until a threshold is reached or packet loss is detected
- congestion avoidance - once a threshold is reached, TCP increases the congestion window linearly to probe for additional bandwidth
- fast retransmist and fast recovery - if packet loss is detected via duplicate acks, then it performs a fast retransmit
- Sequence number wrapping - TCP must handle the sequence number wrapping around
7. Nagle's Algorithm
7. Nagle's Algorithm
- 40ms delay in TCP
- Basically a buffer
- Note that TCP defaults haven't really changed!
- Set
TCP_NODELAYto prevent it
8. Route Map
8. Route Map
- bgp route maps: www.routeviews.org/routeviews/index.php/map/
9. East-West Traffic vs North-South Traffic
9. East-West Traffic vs North-South Traffic
- North south is user facing traffic, east-west is clos. We want east-west at hyperscaler scale because we generate significant fanout, and the oversubscription in north/south is too high if we have one critical spine.
-
The way this works is that each tor advertises its own ASN, and the spine switches usually share an ASN within a single pod. This way, the paths are always leaf -> spine -> leaf -> spine -> server
- Check RFC 7938: Use of BGP for Routing in Large-Scale Data Centers