This article explain fabric drops on the MX10K devices.We will be seeing latency issues around 400G on the LAG interfaces. And this would lead to fabric drops.
Farbic drops would be noticed when traffic exceed 400G on the interface.
labroot@testing> show interfaces vt-0/0/0.5984 extensive
Logical interface vt-0/0/0.5984 (Index 380) (SNMP ifIndex 602) (Generation 201)
Flags: Up Point-To-Point 0x4000 Encapsulation: Virtual-loopback-tunnel
Traffic statistics:
Input bytes : 93327691610835
Output bytes : 0
Input packets: 154898926820
Output packets: 0
Transit statistics:
Input bytes : 93327691610835 415748425352 bps. >>>>> more than 400G
Output bytes : 0 0 bps
Input packets: 154898926820 51864823 pps
Output packets: 0 0 pps
Protocol inet, MTU: Unlimited
Max nh cache: 0, New hold nh limit: 0, Curr nh cnt: 0, Curr new hold cnt: 0, NH drop cnt: 0
Generation: 268, Route table: 13
Flags: Sendbcast-pkt-to-re
Protocol mpls, MTU: Unlimited, Maximum labels: 3, Generation: 271, Route table: 13
Flags: Is-Primary
labroot@testing> show pfe statistics traffic | match fabric
Fabric Input : 331400532316 215679852 pps
Fabric Output : 331370310308 215698085 pps
Fabric drops : 3373634890 >>>>>>fabric drop noticed.
Fabric Input : 332694844651 215730135 pps
Fabric Output : 332664652658 215745895 pps
Fabric drops : 3436072900
Packet Forwarding Engine on MX10003 is based on the third generation of the Trio chipset architecture---namely, the EA (Eagle) ASIC. The Packet Forwarding Engine offers 400 Gbps of WAN and fabric bandwidth each.
Since the single PFE instance has the capacity of only 400G the PFE instance hosting the interface is getting oversubscribed leading to the pressure and the fabric drops reported by multiple source PFEs. This is expected behaviour as the fabric traffic being sent is beyond EA's 400G capacity.