bridge out: extending rfc 2544 for dcb devices timmons c. player david newman ietf bmwg interim...
TRANSCRIPT
Bridge Out:Extending RFC 2544 for DCB
Devices
Timmons C. PlayerDavid Newman
IETF BMWG interim meeting, 30 October 2009
Agenda
• World’s shortest DCB intro• Limitations of throughput for DCB• Limitations of latency for DCB• Other problems• New metrics for DCB testing
Introducing DCB
• DCB (aka DCE, CEE) converges data, storage onto single network
• IEEE 802.1Qbb (aka PFC) adds flow control per VLAN priority
• Other DCB mechanisms for: – Capabilities exchange (DCBX)– Congestion notification (802.1Qau)– Shaping (802.1Qaz)
What’s wrong with throughput?
• RFC 1242 throughput is fine – for Ethernet– Canonical method: Measure oload with 0
loss, followed by oload with packet loss– Highest zero-drop rate is the throughput
rate
• This does not work for DCB
What’s wrong with throughput?
• Loss should never occur with DCB– Flow control throttles transmitters– Impossible to have success case, then fail
case• No distinction between iload and oload– Device that forwards 0 packets could have
“line-rate throughput” in DCB context• No distinction among traffic classes– Different classes may (and probably will)
have different maximum forwarding rates
What’s wrong with latency?
• RFC 2544, section 26.2, requires measurement at throughput rate– Oops: There is no throughput rate
• RFC 1242 uses different measurements for store-and-forward, bit-forwarding– Oops: DCB devices may alternate modes
• RFC 2544 does not measure per class
What else can go wrong?
• 2544/2889 tests use “lock step” pattern
• 1 -> [2,3,4]; 2 -> [3,4,1]; 3-> [4,1,2]; etc.
• Very regular packet departure intervals
What else can go wrong?
• DCB devices quickly go out of lock step• Not just per-port but also per-class• Much tougher on schedulers
Traffic classXOFF/XON
interval (µsec)Inter-PFC burst interval
(µsec)
P1 200 500
P2 150 450
P3 300 700
DCB testing: What’s new
• Proposed new work item:http://www.ietf.org/id/draft-player-dcb-benchmarking-00.txt
• New metric: Queueput– Measures MOL per classification– Multiple queueputs, one per classification,
are possible
• Maximum forwarding rate– Same concept as in 2285/2889– For DCB, more meaningful than throughput– Extended to measure per classification
DCB testing: What’s new
• Back-off measures DUT PFC overhead– Conceptually similar to 2544 frame loss
test– Offer traffic above queueput rate; then
reduce iload until the DUT no longer pauses ingress traffic
–Measure per classification
DCB testing: What’s new
• Back-to-back– Conceptually similar to back-off in RFCs
1242/2544– Extended to measure per classification
• Other DCB metrics?