Tutorial Survey of LL-FC Methods for Datacenter Ethernet 101 Flow Control M. Gusat Contributors: Ton Engbersen, Cyriel Minkenberg, Ronald Luijten and Clark Jeffries 26 Sept. 2006 IBM Zurich Research Lab 1 Outline • Part I Requirements of datacenter link-level flow control (LL-FC) Brief survey of top 3 LL-FC methods PAUSE, aka. On/Off grants credit rate Baseline performance evaluation • Part II Selectivity and scope of LL-FC per-what? : LL-FC’s resolution 2 Req’ts of .3x’: Next Generation of Ethernet Flow Control for Datacenters 1. Lossless operation No-drop expectation of datacenter apps (storage, IPC) Low latency 2. Selective Discrimination granularity: link, prio/VL, VLAN, VC, flow...? Scope: Backpressure upstream one hop, k-hops, e2e...? 3. Simple... PAUSE-compatible !! 3 Generic LL-FC System RTT • One link with 2 adjacent buffers: TX (SRC) and RX (DST) • Round trip time (RTT) per link is system’s time constant LL-FC issues: link traversal (channel Bw allocation) RX buffer allocation pairwise-communication between channel’s terminations signaling overhead (PAUSE, credit, rate commands) backpressure (BP): increase / decrease injections stop and restart protocol 4 FC-Basics: PAUSE (On/Off Grants) Stop Go TX Queues “Over -run”= Send STOP FC Return path RX Buffer OQ Data Link Threshold PAUSE BP Semantics : STOP / GO / STOP.. Xbar Down-stream Links * Note: Selectivity and granularity of FC domains are not considered here. 5 FC-Basics: Credits Xbar * Note: Selectivity and granularity of FC domains are not considered here. 6 Correctness: Min. Memory for “No Drop” "Minimum“: to operate lossless => O(RTTlink) – Credit : 1 credit = 1 memory location – Grant : 5 (=RTT+1) memory locations Credits – Under full load the single credit is constantly looping between RX and TX RTT=4 => max. performance = f(up-link utilisation) = 25% Grants – Determined by slow restart: if last packet has left the RX queue, it takes an RTT until the next packet arrives 7 PAUSE vs. Credit @ M = RTT+1 "Equivalent" = ‘fair’ comparison 1. Credit scheme: 5 credit = 5 memory locations 2. Grant scheme: 5 (=RTT+1) memory locations Performance loss for PAUSE/Grants is due to lack of underflow protection, because if M < 2*RTT the link is not work-conserving (pipeline bubbles on restart) For equivalent (to credit) performance, M=9 is required for PAUSE. 8 FC-Basics: Rate • • • RX queue Qi=1 (full capacity). Max. flow (input arrivals) during one timestep (Dt = 1) is 1/8. Goal: update the TX probability Ti from any sending node during the time interval [t, t+1) to obtain the new Ti applied during the time interval [t+1, t+2). • Algorithm for obtaining Ti(t+1) from Ti(t) ... => • Initially the offered rate from source0 was set = .100 , and from source1 = .025. All other processing rates were .125. Hence all queues show low occupancy. • At timestep 20, the flow rate to the sink was reduced to .050 => causing a congestion level in Queue2 of .125/.050 = 2.5 times processing capacity. • Results: The average queue occupancies are .23 to .25, except Q3 = .13. The source flows are treated about equally and their long-term sum is about .050 (optimal). 9 Conclusion Part I: Which Scheme is “Better”? • PAUSE + simple + scalable (lower overhead of signalling) - 2xM size required • Credits (absolute or incremental) + are always lossless, independent of the RTT and memory size + adopted by virtually all modern ICTNs (IBA, PCIe, FC, HT, ...) not trivial for buffer-sharing protocol reliability scalability • At equal M = RTT, credits show 30+% higher Tput vs. PAUSE *Note: Stability of both was formally proven here • Rate: in-between PAUSE and credits + adopted in adapters + potential good match for BCN (e2e CM) - complexity (cheap fast bridges) 10 Part II: Selectivity and Scope of LL-FC “Per-Prio/VL PAUSE” The FC-ed ‘link’ could be a physical channel (e.g. 802.3x) virtual lane (VL, e.g. IBA 2-16 VLs) virtual channel (VC, larger figure) ... • Per-Prio/VL PAUSE is the often proposed PAUSE v2.0 ... • Yet, is it good enough for the next decade of datacenter Ethernet? • Evaluation of IBA vs. PCIe/As vs. NextGen-Bridge (PrizmaCi) 11 Already Implemented in IBA (and other ICTNs...) • IBA has 15 FC-ed VLs for QoS • However, IBA doesn’t have VOQ-selective LL-FC • “selective” = per switch (virtual) output port So what? • SL-to-VL mapping is performed per hop, according to capabilities Hogging - aka buffer monopolization, HOL1-blocking, output queue lockup, single-stage congestion, saturation tree(k=0) How can we prove that hogging really occurs in IBA? A. Back-of-the-envelope reasoning B. Analytical modeling of stability and work-conservation C. Comparative simulations: IBA, PCI-AS etc. (next slides) (papers available) 12 IBA SE Hogging Scenario • Simulation: parallel backup to a RAID across an IBA switch TX / SRC 16 independent IBA sources, e.g. 16 “producer” CPU/threads SRC behavior: greedy, using any communication model (UD) SL: BE service discipline on a single VL – (the other VLs suffer of their own ) Fabrics (single stage) 16x16 IBA generic SE 16x16 PCI-AS switch 16x16 Prizma CI switch RX / DST 16 HDD “consumers” t0 : initially each HDD sinks data at full 1x (100%) tsim : during simulation HDD[0] enters thermal recalibration or sector remapping; consequently » HDD[0] progressively slows down its incoming link throughput: 90, 80,..., 10% 13 First: Friendly Bernoulli Traffic 2 Sources (A, B) sending @ (12x + 4x) to 16*1x End Nodes (C..R) Fig. from IBA Spec R achievable performance Throughput loss aggregate throughput link 0 throughput reduction 14 Myths and Fallacies about Hogging • • Isn’t IBA’s static rate control sufficient? No, because it is STATIC • • IBA’s VLs are sufficient...?! No. VLs and ports are orthogonal dimensions of LL-FC • • 1. VLs are for SL and QoS => VLs are assigned to prios, not ports! 2. Max. no. of VLs = 15 << max (SE_degree x SL) = 4K Can the SE buffer partitioning solve hogging, blocking and sat_trees, at least in single SE systems? No. 1. Partitioning makes sense only w/ Status-based FC (per bridge output port - see PCIe/AS SBFC); IBA doesn’t have a native Status-based FC 2. Sizing becomes the issue => we need dedication per I and O ports M = O( SL * max{RTT, MTU} * N2 ) very large number! Academic papers and theoretical disertations prove stability and workconservation, but the amounts of required M are large 15 Conclusion Part II: Selectivity and Scope of LL-FC Despite 16 VLs, IBA/DCE is exposed to the “transistor effect”: any single flow can modulate the aggregate Tput of all the others Hogging (HOL1-blocking) requires a solution even for the smallest IBA/DCE system (single hop) Prios/VL and VOQ/VC are 2 orthogonal dimensions of LL-FC Q: QoS violation as price of ‘non-blocking’ LL-FC? • Possible granularities of LL-FC queuing domains: A. CM can serve in single hop fabrics also as LL-FC B. Introduce VOQ-FC: intermediate coarser grain no. VCs = max{VOQ} * max{VL} = 64..4096 x 2..16 <= 64K VCs Alternative: 802.1p (map prios to 8 VLs) + .1q (map VLANs to 4K VCs)? Was proposed in 802.3ar... 16 Backup 17 LL-FC Between Two Bridges Switch[k] Switch[k+1] TX Port[k,j] RX Port[k+1, i] VOQ[1] RX Mgnt. Unit (Buffer Allocation) VOQ[n] TX Scheduler "send packet" RX Buffer LL-FC TX Unit LL-FC Reception “return path of LL-FC token" 18