Be Beyond nd Jai ain’s n’s Fai Fairne ness Inde ndex: : Set Settin ing The e Bar ar For or the e Dep eploy loymen ent of of Con ongest estion ion Con ontrol rol Alg lgorit orithms Ranysha Wa Ware Matthew K. Mu Ma Mukerjee Justi Ju stine S e Sher erry Sr Srinivasan Se Seshan Carnegie Mellon Nefeli Carnegie Mellon Carnegie Mellon University Networks University University 1
: ! I I hav have e desi designed gned a a new new CCA: w ! is Ho How do do we we sho show is reason onable le to to deploy in th the Inte ternet? t? 2
fairness to show that ! is reasonably deployable We typically use fairnes alongside " , a legacy algorithm. 3
But ev every eryone one falls falls short hort of achieving fair outcomes. 4
But ev every eryone one falls falls short hort of achieving fair outcomes. Cubic can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” 5
But ev every eryone one falls falls short hort of achieving fair outcomes. CUBIC can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” BBRv1 can be unfair to Cubic, but “we are looking at modeling shallow buffer situations”. 6
But ev every eryone one falls falls short hort of achieving fair outcomes. CUBIC can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” BBRv1 can be unfair to Cubic, but “we are looking at modeling shallow buffer situations”. PCC Vivace can be unfair to Cubic, but “as the number of CUBIC senders increases, it achieves the best fairness among new generation protocols.” 7
But ev every eryone one falls falls short hort of achieving fair outcomes. CUBIC can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” BBRv1 can be unfair to Cubic, but “we are looking at modeling shallow buffer situations”. PCC Vivace can be unfair to Cubic, but “as the number of CUBIC senders increases, it achieves the best fairness among new generation protocols.” Copa can be unfair to Cubic, but “is much fairer than BBR and PCC” and “uses bandwidth Cubic does not utilize.” 8
Ev Everyone makes excu excuses ses wh why th thei eir r algori rith thm is s sti still rea reason sonable le to to dep eploy loy desp espite ite unf unfair out utcomes. 9
Th This is talk lk: We We need need a a pr pract actical cal depl deployment ent th threshold: a a bo bound und on n ho how w e ! , aggr aggressi essive , a a new new CCA, , can can be to " , be , the he st stat atus us quo quo. 10 10
Ou Outline: 1. 1. What are desir irable le prop opertie ies of of a deploy loyment threshold old? 2. 2. We defin ine a new deploy loyment th threshold: harm. 11 11
Ou Outline: 1. 1. What are desir irable le prop opertie ies of of a deploy loyment threshold old? 2. 2. We defin ine a new deploy loyment th threshold: harm. 12 12
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEM DEMAN AND- METR ME TRIC AW AWAR ARE STATU STA TUS-QU QUO FUTURE- FU PR PRACT CTICA CAL BIAS BI ASED ED PR PROOF 13 13
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEMAN DEM AND- METR ME TRIC AWAR AW ARE STA STATU TUS-QU QUO FUTURE- FU PRACT PR CTICA CAL BI BIAS ASED ED PR PROOF 14 14
A deployment threshold needs to be pract practical: ical: should be feasible for new CCA to meet threshold. 15 15
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEMAN DEM AND- METR ME TRIC AWAR AW ARE STA STATU TUS-QU QUO FUTURE- FU PRACT PR CTICA CAL BI BIAS ASED ED PR PROOF 16 16
Slow bottleneck link 17 17
Slow bottleneck link CCA: ! CCA 18 18
La Latenc ncy: 5 ms Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA: ! CCA 19 19
La Latenc ncy: 5 ms Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA: " CCA CCA: ! CCA 20 20
La Latenc ncy: 5 ms Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA: " CCA CCA: ! Do Downl nload d spe peed: d: 5 Mbps CCA 21 21
Latenc La ncy: 5 ms 100 100 ms Li Link nk capa pacity: 10 Mbps Downl Do nload d spe peed: d: 5 Mbps CCA: " CCA CCA: ! Do Downl nload d spe peed: d: 5 Mbps CCA 22 22
A deployment threshold needs to be mu multi lti-me metr tric ic : can account for performance metrics beyond just throughput. Latenc La ncy: 5 ms 100 100 ms Link Li nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA: " CCA CCA: ! Downl Do nload d spe peed: d: 5 Mbps CCA 23 23
Me Metrics l like l latency c y cannot b be “d “divided fa fairl rly”. ”. 24 24
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEMAN DEM AND- METR ME TRIC AWAR AW ARE STA STATU TUS-QU QUO FUTURE- FU PRACT PR CTICA CAL BI BIAS ASED ED PR PROOF 25 25
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps CCA: ! CCA 26 26
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps CCA: ! CCA 27 27
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps CCA: " CCA CCA: ! CCA 28 28
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps 9 Mbps CCA: " CCA CCA: ! Downl Do nload d spe peed: d: 1 1 Mb Mbps CCA 29 29
A deployment threshold needs to be st status-quo quo bias biased ed : based only on impact of ! on " , not vice-versa. d: 10 Mbps 9 Mbps Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: CCA: " CCA CCA: ! Do Downl nload d spe peed: d: 1 1 Mb Mbps CCA 30 30
Ja Jain’s f fairness i index i x is n not s status- quo quo bi biased. ased. 31 31
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEMAN DEM AND- METR ME TRIC AWAR AW ARE STA STATU TUS-QU QUO FUTURE- FU PRACT PR CTICA CAL BI BIAS ASED ED PR PROOF 32 32
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA: ! CCA 33 33
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA: ! CCA 34 34
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA: " CCA CCA: ! CCA 35 35
Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA: " CCA CCA: ! Do Downl nload d spe peed: d: 7 Mbps CCA 36 36
A deployment threshold needs to be dem demand and-aware aware: : do not penalize ! when " has inherently poor performance. Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA: " CCA CCA: ! Do Downl nload d spe peed: d: 7 Mbps CCA 37 37
Ma Max-mi min f fairness i is d dema mand a aware, equal equal-ra rate fa fairn rness is not. 38 38
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEMAN DEM AND- METR ME TRIC AWAR AW ARE STA STATU TUS-QU QUO FUTURE- FU PRACT PR CTICA CAL BI BIAS ASED ED PR PROOF 39 39
A deployment threshold needs to be future future-proof proof : useful on a future Internet where none of today’s current CCAs are deployed. 40 40
A deployment threshold needs to be future future-proof proof : useful on a future Internet where none of today’s current CCAs are deployed. Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 1 Mbps CCA: ! CCA 41 41
A deployment threshold needs to be future future-proof proof : useful on a future Internet where none of today’s current CCAs are deployed. Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA: ! CCA 42 42
Does ! need to be nice to " and # or just " ? Li Link nk capa pacity: 10 Mbps Downl Do nload d spe peed: d: 5 Mbps CCA: " CCA CCA: ! CCA 43 43
A future-proof threshold would only require ! to be nice to " Li Link nk capa pacity: 10 Mbps Downl Do nload d spe peed: d: 5 Mbps CCA: " CCA CCA: ! CCA 44 44
TC TCP-fri friendliness is not fu future re-pr proof. 45 45
We identify 5 des 5 desirable properties irable properties for a deployment threshold. MULTI MU TI- DEMAN DEM AND- METR ME TRIC AWAR AW ARE STA STATU TUS-QU QUO FUTURE- FU PRACT PR CTICA CAL BI BIAS ASED ED PR PROOF 46 46
Ou Outline: 1. 1. What are desir irable le prop opertie ies of of a deploy loyment threshold old? 2. 2. We defin ine a new deploy loyment th threshold: harm. 47 47
When showing deployability: we run experiments of ! vs. " and me measure p perfor orma mance . 48 48
Recommend
More recommend
