3 ms·

Consider the volume of traffic going over big backbones though (e. g. 100gbps per second). Let's say one of the devices terminating this backbone has a piece o

by hueving 10y ago

Consider the volume of traffic going over big backbones though (e. g. 100gbps per second). Let's say one of the devices terminating this backbone has a piece of memory go bad that mangles IP payloads. At 8.3 Mpps (1500byte mtu line rate), a 1 in 4 billion chance would trigger on average every 9 minutes or so. That only assumes random probably with checksum collisions, but its interesting to see how the volumes of traffic moved around the Internet now have brought 'rare' events into much more frequency.

phicoh 10y ago
The ethernet CRC is designed for packets on the wire. On the wire you have a bit error rate and sometimes burst errors. For bit errors, as long as no packet has 32 bit errors the CRC will always catch them. For burst errors, if the error sets all bits to zero (or to one) then the CRC will catch it. This is different from random memory corruption. Note that while a packet is the memory of a router you typically only have to TCP checksum to protect the packet, which is a rather weak 16-bit sum. So it is safe to assume that TCP will fail to detect many failures and if you care add a sha2 hash. Why sha2, because while md5 is perfectly fine for random errors, we should should make sure to kill all use of insecure hash functions. Otherwise they keep popping up in contexts where errors are not random.
kingosticks 10y ago
Current generation Cisco and Nokia core routers seem to be around 400Gbps per slot. Juniper is claiming something wildly higher (up to ~1.5Tbps) so I guess they are ahead but double counting like some of these guys do. But yes, at these speeds, definitely more common, and getting more frequent all the time. http://www.juniper.net/techpubs/en_US/release-independent/junos/topics/reference/general/fpc-ptx5000-supported.html http://www.juniper.net/techpubs/en_US/release-independent/ju... EDIT: changed 3Tbps to 1.5Tbps.
wang_li 10y ago
>Consider the volume of traffic going over big backbones though (e. g. 100gbps per second). Let's say one of the devices terminating this backbone has a piece of memory go bad that mangles IP payloads. At 8.3 Mpps (1500byte mtu line rate), a 1 in 4 billion chance would trigger on average every 9 minutes or so. I think someone would notice when only one packet per nine minutes successfully traverses their 100 Gbps core switch. So mission accomplished. :)

by hueving 10y ago

phicoh 10y ago

The ethernet CRC is designed for packets on the wire. On the wire you have a bit error rate and sometimes burst errors. For bit errors, as long as no packet has 32 bit errors the CRC will always catch them. For burst errors, if the error sets all bits to zero (or to one) then the CRC will catch it. This is different from random memory corruption. Note that while a packet is the memory of a router you typically only have to TCP checksum to protect the packet, which is a rather weak 16-bit sum. So it is safe to assume that TCP will fail to detect many failures and if you care add a sha2 hash. Why sha2, because while md5 is perfectly fine for random errors, we should should make sure to kill all use of insecure hash functions. Otherwise they keep popping up in contexts where errors are not random.

kingosticks 10y ago

Current generation Cisco and Nokia core routers seem to be around 400Gbps per slot. Juniper is claiming something wildly higher (up to ~1.5Tbps) so I guess they are ahead but double counting like some of these guys do. But yes, at these speeds, definitely more common, and getting more frequent all the time. http://www.juniper.net/techpubs/en_US/release-independent/junos/topics/reference/general/fpc-ptx5000-supported.html http://www.juniper.net/techpubs/en_US/release-independent/ju... EDIT: changed 3Tbps to 1.5Tbps.

wang_li 10y ago

>Consider the volume of traffic going over big backbones though (e. g. 100gbps per second). Let's say one of the devices terminating this backbone has a piece of memory go bad that mangles IP payloads. At 8.3 Mpps (1500byte mtu line rate), a 1 in 4 billion chance would trigger on average every 9 minutes or so. I think someone would notice when only one packet per nine minutes successfully traverses their 100 Gbps core switch. So mission accomplished. :)