David Ahern

@davem_dokebi @SHRIJEET2 has a perspective with comments on what it means for network design: blog.enfabrica.net/ultra-ethernet…

What do AI folks need other than RDMA a D direct data placement? Asking for a friend...

@atoonk So many questions… start with, AF_XDP on both ends ( post says packet generator)?

Some fun with Go this weekend! A packet generator ala Iperf3, but using AF_XDP. getting 186Gbs 🤓🚀🤩 This is *a lot* faster than what I see with Iperf3

We are pleased to announce that we have joined the Ultra Ethernet Consortium (UEC)! Come swing by our booth #586 at #SuperComputing to learn more! bit.ly/47AYJ8G

CXL a new beggining according to @SHRIJEET2 @netdev01 0x17 #netdevconf

Multiple round trips across the mem + PCI bus for data from/to GPU/SSD/Network are killing us in ML. Time to change how Linux handles it say Mina Almasry et al in this @netdev01 0x17 talk. Exciting times in Linux and the networking world!!!😹 netdevconf.info/0x17/5 #netdevconf

How does CXL disaggregation of memory and compute play with networking? Put buffer and control state placement next to offload devices +IO and then merge IO + Compute expansion protocols. @SHRIJEET2 will leads the the CXL and networking BoF @netdev01 0x17. Join us!#netdevconf

So your NIC hardware is not ready but you need to write the driver. Things that make you go hmmm.. 🤔...... Come and learn from @dsahern @netdev01 0x17 tutorial on how to deal with this netdevconf.info/0x17/sessions/… #netdevconf

Courtesy of @Nasdaq - a live shot of Times Square just this past Friday! To all our incredible employees, partners and investors, thank you for your role in getting @enfabrica_ to where we are today. We couldn't do it without you and this is just the beginning!

@NolanSIGBUS @networkservice @Ra1t0_Bezar1us Interesting, the repo was removed from github. I should have a clone somewhere...

@dsahern @networkservice @Ra1t0_Bezar1us Sadly the link to the clsw code in Cumulus's net-next github repo has bit-rotted.

none of the good stuff has switchdev support

@networkservice @NolanSIGBUS @Ra1t0_Bezar1us BCM with switchdev - would love to see it. Having done something similar [1] to only a POC level I will say a $500k contract depends on the details. :-) [1] github.com/dsahern/media/…

@NolanSIGBUS @Ra1t0_Bezar1us upstreaming! hah 🤣 would never dream it. I'd prefer not to only have SAI and it's...progress...be my only window to the merchant silicon world. but it'll do for now.

@vince2_ iperf3 and VM to VM would be 4 memcpy’s , so 16G sounds right.

This Thursday, at netdev 0x16, I talk about how RDMA concepts can be used with socket based networking to achieve most of the properties needed to scale up to high flow rates (netdevconf.info/0x16/session.h…).

Last month at LPC,I talked about the scalability of the Linux networking stack and what S/W needs for a single flow to scale up as line rates increase (lpc.events/event/16/contr…).

You want high performance application networking? fuggedaboutit if you insist on using BSD socket API. @dsahern and @SHRIJEET2 have been in pursuit of perf happyness and in this @netdev01 0x16 they discuss how we get to the promised land #netdevconf netdevconf.info/0x16/session.h…

@blakedot_ @majek04 @networkservice There are a lot of permutations emphasizing how hard it is to cover all possible use cases. More tests can be added to cover your companies' needs so future changes do not cause breakage which takes years to discover.

@blakedot_ @majek04 @networkservice Test cases are key. In 2018 my test suite was not in the kernel tree which is the back and forth on that set; it is now[1]. Some of those tests include address binding with and without VRF. [1] #n3474" target="_blank" rel="nofollow noopener">git.kernel.org/pub/scm/linux/…

On Linux every complex routing problem can supposedly be solved by VRF but nobody ever saw VRF actually working.

@davekempe @toprankinrez @ValdikSS @majek04 I do not recall that OSS talk being recorded.

@ValdikSS @toprankinrez @majek04 Linux VRF is implemented as policy routing. As such you must have a default route in the VRF table (your unreachable entry), and the fib rules need to be properly ordered (local rule moved below l3mdev). All of that is in the slide deck as well as troubleshooting lookups via perf

@toprankinrez @dsahern @majek04 Add IPv6 address and a route to eth0, but not on int0/vrf. Run ping -I vrf0 ya ru Result: PING ya ru(ya ru (2a02:6b8::2:242)) from 2000::1 vrf0 It tries to ping by IPv6 using eth0's source address. Run ip -6 route add unreachable default vrf vrf0 Result: no issues, ping by v4.

@toprankinrez @ValdikSS @majek04 That URL is to Open Source Summit, North America, Sept 2017. The original Cumulus blog post does seem to have vanished after the nvidia acquisition.

@dsahern @ValdikSS @majek04 Thanks for this! I had a link to it on the cumulus site before but they all died.

@ValdikSS @majek04 static.sched.com/hosted_files/o…, slide 44 and 45.

@majek04 I! I configured remote site with gretap inside VRF two days ago. It's very convenient when you need to use different links but don't need it inside main routing table amd don't want to configure policy routing. The drawback is that 'ip r' still shows main r.t. in 'ip vrf exec'.