Scientific best-practices for recurring problems in computer security R& D

Scientific best-practices for recurring problems in computer security R& D

Cad49e6ffc6048dc9c53c77a907632dc?s=128

Daniel Bilar

April 03, 2014
Tweet

Transcript

  1. Scien&fic  Best-­‐Prac&ces  for  Recurring   Problems  in  Computer  Security  R

     &  D   Daniel  Bilar   Director  of  Research,  Siege  Technologies     Manchester,  New  Hampshire,  USA   dbilar@acm.org   @daniel_bilar     SyScan  2014   Singapore   April  3rd,  2014   1  
  2. •  Dynamic  SyScan  update:  Car  CAN  R  &  D  

    •  Signals/  Side  Channels   •  Side  Channels  &  Methods   –  Detec&on   –  Representa&on   –  Analysis   •  Case  studies   –  ROPe,  Bochspwn,  Cyber  Mission  Planning   •  Epilogue/Blueprint     Outline  of  Talk   2   Thanks  for  help/ inspira&on  and   apprecia&on  to     Thomas  Dullien   Ero  Carrera   Alex  So-rov   Travis  Goodspeed   Anna  Shubina     Sergey  Bratus   Rebecca  Shapiro   Jason  Geffner   Jon  Stuart   Georg  Wicherski   Mateusz  Jurczyk   Gynvael  Coldwind     +  many  more    
  3. Addendum:  Car  CAN  bus  hacking   •  Did  a  subset

     of  what  Chris  and  Charlie  did  in  March  2013,   presented  at  RECon  June  2013     •  “Hot-Wiring of the Future”   –  Lots  of  &ps,  (free,  $)  tools,  workflow/methodology,  Costs:  $6k  (2   cars)   •  Sponsored  3  undergraduate  students  (knew  nothing  at  all   about  RE)  who  learned  how  to  reverse,  hook  up  boards,   use  goodThopter  and  denial  of  view/manipula&on  of  CAN   dashboard  in  3  months     3   GoodThopter10   CAN  Bus   So`ware  Package   User   hap://&nyurl.com/CarCAN2013  
  4. Work Accomplished - Methodology   Confirm  Inferences,   Test  Responsiveness

      Boundary  Analysis   Genera&ve  Fuzzing   Confirm  New   Inferences   4! ArbID   Fuzzing  Response   Refined  Inferences   513   Dashboard  components  change   Bytes  0,  1  =  RPM   Byte  4  =  Speedometer   A Case Study – ID 513   engineering.dartmouth.edu 7 March 2013
  5. 0   20   40   60   80  

    100   120   0   20   40   60   80   100   120   Speedometer  Reading  (MPH)   Data  Byte  4  Value   5! Higher Level Protocol: ID 513   engineering.dartmouth.edu 7 March 2013
  6. Higher Level Protocol: ID 1056   1056   DLC  =

     8   Engine   Temp   Odometer   Ba[ery   Charge   Engine  Clock   Dashboard   Warnings   Check  Fuel   Cap   Unused   Counter   S O F   Arbitra&on  ID   Control  Field   Data  Field   E O F   6! engineering.dartmouth.edu 7 March 2013
  7. Signals:  Side  Channel   •  Side  channels  =   observables

     that  are   emiaed  by  ac&ve  systems   •  On  a  computer  system,   these  can  be  observed   &me,  power,  OS  events,   EM  radia&on,  characteris&c   acous&c  spectral  signature   and  more   •  2014:  Mark  Stoefnger’s   NTU  group  is  doing  cufng   edge  punctuated  hardware   magne&c  field  side  channel   analysis  (SCA)   Graph  from  [Hund2013].  Generic  &ming  side   channel  aaack  against  MMU  system  to  infer   informa&on  about  the  privileged  address  space   layout     Some  innova&ve  aaacks:  data  structures  (2007   &ming  aaacks  against  databases),  protocols  and   underlying  algorithms  (2007  QoS  aaacks  against   balancing  algorithms,  MMU  (cache)  and  more   7  
  8. Side  Channels  as  Consilient  Evidence   •  Generalize:  Use  side

     channel  evidence  to  reason  about   system/sub-­‐system  (internal  states)   •  Whewell’s  “Consilience  of  Induc&on”   –  Concept  of  aggregate  evidence   –  Convergence  of  several,  ideally  independent  hypotheses  serves  to   strengthen  conclusion   •  Ques&on:  What  side  channels  are  available,  easy  to  access,   analyze,  expressive,  type  I/II  error  etc   8  
  9. Three  Ques&ons  on  Signals   •  Want  ac&onable  handling  of

     higher  dimensional   signal  dynamics  as  they  occur  in  live  computer   systems  as  side  channels   –  Ac&onable  is  to  be  understood  as  useful  in  prac&ce   –  Higher  dimensional  refers  to  six  or  more  data   dimensions     –  Dynamics  emphasizes  the  signal’s  unfolding  in   temporal  and  feature  space   •  We’ll  discuss   –  How  to  represent  such  signals     –  How  to  detect  signals  with  various  characteris&cs     –  How  to  prevent  or  mi&gate  the  leaking  of  such  signals     9  
  10. Signal  Representa&on:  Visuals   10  

  11. Signal  Representa&on:  Visuals  &  more   •  Fundamental  cogni&ve  limits

     for  visuals  very  hard  to  overcome  for   high  dimensional  representa&on   •  See  Starlight  project,  NIDS  graphs,  Edward  Tu`e  book  series,   Marty  “  Applied  Security  Visualiza&on”,  Paley  “Visual  Analy&cs”   Rich  gamut  of  of  human  senses  remain  neglected   •  Aural  (hearing),  hap&c  (touch),  ves&bular  (balance  and   accelera&on),  kinesthe&c,  thermocep&on  (temperature),  etc   11  
  12. Maximal  Informa&on-­‐based   Nonparametric  Explora&on   (MINE)  sta&s&cs   § 

    General:  Captures  wide   range  of  associa-ons   between  pairs  of  variables   (linear,  exponen&al,   periodic,  non-­‐func&ons)   §  Equitable:  Assigns  similar   scores  to  equally  noisy   rela&onships  of  different   types  [Reshef2011sup]   Signal  Detec&on:  MINE  Sta&s&cs   Table  from  [Reshef2011]   12   MIC  captures  general  rela-onship  strength   MIC-­‐r^2  captures  non-­‐linearity  (Not  shown)     MCN  captures  complexity   MAS  captures  departure  from  monotonicity   MEV  captures  closeness  to  being  a  func-on  
  13. Signal  Detec&on:  Associa&on  ‘Types’   13   Graphs  from  [Reshef2011sup]

     
  14. Signal  Preven&on:  Side  Channel  Leaks     •  Possible  to

     control  rate  but  not  eliminate  side  channels   •  Recently,  Goldwasser  (MIT/Technion)    and  Rothblum  offer   a  prac&cal  way  forward     •  Resis&ng  leakage  at  design  (me  and  offers  progress   towards  formula(on  of  automa(c  approaches  that   generate  “leakage-­‐resilience”  programs  for  a  wide  range  of   side  channel  aaacks     •  Proved  that  for  any  computa&onally  unbounded  A   observing  the  results  of  computa&onally  unbounded   leakage  func&ons,  will  learn  no  more  from  its  observa&ons   than  it  could  given  blackbox  access  only  to  the  input-­‐output   behavior  of  P     •  Result  is  uncondi(onal  and  does  not  rely  on  any  secure   hardware  components   14  
  15. Cyber  Mission  Planning   •  Cyber-­‐opera&ons  have  poten&al  to  be

     more   pinpointed  than  kine&c  counterpart     – Minimize  collateral  damage  by  ‘crisp’  targeted   opera&ons     •  However,  unlike  kine&c  planning  (centuries  of  well-­‐ understood  natural  laws),  cyber-­‐planning  lacks   founda-onal  corpus  of  predic-ve  laws     “Natural  Laws  for/of/in  Virtual  Reality”     15  
  16. Thread  /  Process/  Cache  Execu&on   •  Behavior  arises  as

     a  complex   interac&on  of  &ming  of  memory   requests  (program  behavior),   cache  coherence  protocol   (dependent  on  MA),  or  thread  pre-­‐ emp&on  (depending  on  OS)   [Alistarh2014]   •  Modern  opera&ng  systems  (OS)   and  microarchitectures  (MA)  =   dynamic  complex  feedback  system   that  tries  to  con&nuously  minimize   CPI  (cycles  per  instruc&on)   –  Memory  latency  is  boaleneck,   hence  memory  hierarchies  from  ns   to  s   •  OS  and  MA  con-nuously  solve  a   -me-­‐space  op-miza-on  problem   to  ‘fla[en’  (parallelize)  sequen-al   processing   16   OS  schedules  thread (s),  positions  data  in   memory     (space/time   optimization)   Microarchitectures  atomizes   and  interleaves  thread  ops  to   minimize  CPI   Memory   hierarchy     (access   latency  ns  -­‐   seconds)   Threads Data flow Feedback signals Hardware,  user,  I/O   Interrupts (stochastic) Minimize   expected  data   latency  based   on  feedback   Program(s)    (requires   space-­‐time  from  OS/MA)  
  17. Maximal  Informa&on-­‐based   Nonparametric  Explora&on  (MINE)   sta&s&cs    

    Intui&on:  (Simple)  asset  ‘signals’  are   reflected  in  convex/concave   parabola-­‐type  curves  in  -me     Iden&fy  signals  that  are  less  periodic   (lower  MAS),  less  linear  (MIC-­‐r^2),   but  s&ll  a  func&on  (higher  MEV)     “not  a  heartbeat,  not  a  shoo(ng  star,   but  s(ll  a  func(on”     Signal  Detec&on:  MINE/MIC   MIC  captures  general  rela-onship  strength   MIC-­‐r^2  captures  non-­‐linearity  (Not  shown)     MCN  captures  complexity   MAS  captures  departure  from  monotonicity   MEV  captures  closeness  to  being  a  func-on   17  
  18. Issue:  Experimental  Factors   •  Free  NIST  tools:  Automated  Combinatorial

      Tes&ng  for  So`ware  (ACTS)  and  Coverage   Measurement  (CCM)   •  Cut  down  combinatorial  explosion:  2-­‐way,  3-­‐ way,  n-­‐way  tes&ng   18  
  19. Asset-­‐Target  Matching     •  Say  asset  tested  on  

    configura&on  A  and  it  has  18   categories  (e.g.  language,  OS/ patch,  service  running,   workload,  etc)  @  dozen  of   values   •  How  similar  is  the  unknown   configura&on  B  to  A?   •  Ques&on  of  distance   –  Easy  enough  for  ra&o  data  (like   Kelvin),  much  harder  for   categorical  data  (like  OS  type)   •  Table  shows  14  categorical   distance  (similarity)  measures   •  Differ  primarily  how  they   weigh  matches  and   mismatches  between   categories   ‘Success’  func&ons  are  not  smooth  but  stepped     •  Can  shove  10cm^3  (~elas&c)  object  through   9.8cm^3  hole  –  >  smooth  success   degrada&on   •  Difference  between  patch1  and  patch2  is   difference  between  works/doesn’t  work  -­‐>   step/catastrophic  success  degrada&on   19   Table  from  [Boriah2013]  
  20. Case  study:  ROPe   •  ROPe:  Detec&on  of  kernel-­‐level  ROP

     through  branch  return  mispredic&ons   •  16  (N)  entry  shadow  stack  of  call-­‐  sites  /  return  addresses   –  0x89  –  BR_MISP_EXEC.*:  mispredicted  executed  branches     –  0x800  –  .RETURN_NEAR:  normal,  near  ret     –  0x8000  –  .TAKEN:  uncondi&onal  branch   •  PMC  interrupts  a`er  certain  number  of  mispredic&ons  (N/2  =  8)   –  Upon  interrupt,  handler  checks  MSR  Last  Branch  Recording  (LBR)  whether  targets  of   the  previously  executed  instruc&ons  are  preceded  by  an  instruc&on   –  If  not  -­‐>  likely  ROP  (chain)  induced   20   Started  telling  Travis   Goodspeed  at  RECon   2013  about  this  and   a`er  less  than  8   seconds  he  exclaims   and  I  quote:  Holy   cow!  That’s  a  $^&*   brilliant  idea  once   you  understand  it!    J     Graph  from  G.  Wicherski,  SysCan  2013  
  21. Sugges&ons  for  ROPe   •  Generalize  ROP/0x8889  insight  :  

    Side  channel  ‘spectral  signature’     for  variety  of  interes&ng  aaacks   –  JOP,  ‘weird  machine’-­‐inducers,   hardware-­‐based  aaacks,  …   –  Addi&onal  OS/MA  vents   Workplan  (high-­‐level):   •  Iden&fy  signals  of  interest   –  Scope  with  MINE  [Reshef2011]   •  Signal  periodicity  analysis   –  DSP,  System  Iden&fica&on  tools     •  Scien&fically  valid  experiment   setup   –  Use  procedures  [Mont2012]   21   PMC  measurements  over  &me  for  programs   in  SPEC  benchmark  suite.  Graph  from   [Demme2013].  Good  results  with  4-­‐dim   spectral  signature  from    x86  MA  events   0x0440  -­‐-­‐  L1D_CACHE_LD.E_STATE   0x0324  -­‐-­‐  L2_RQSTS.LOADS   0x03b1  -­‐-­‐  UOPS_EXECUTED.PORT  (1  or  2)   0x7f88  -­‐-­‐  BR_INST_EXEC.ANY  
  22. Case  study:  Bowspwn   •  Study  of  Double  fetch  opera&ons

      –  Two  virtual  address  reads  from  kernel  mode   thread  close  in  &me   –  Virtual  address  concurrently  writable  by  ring-­‐3   threads     •  Assump&on  of  value  consistency  over  &me   gives  rise  to  race  condi&on   –  User  address  space  is  shared  across  ring0  /   ring3     –  User-­‐mode  memory  regions  can  be  modified  at   any  &me  by  concurrent  ring3  thread   •  Bochspwn  idea:  Extend  &me  window   between  value  Check  and  value  Use  to  give   ring3  aaack  tread  opportunity  to  modify   value   –  How?  Bleed  &me  by  slow  cache  line  and  page   boundaries,  non-­‐cacheability,  TLB  flushing   (2500x  slowdown  achieved)   •  69  (!)  LaTex  pages  at  SyScan  2013   –  Fundamental  applied  security  paper,  vital  for   safer  concurrent  programming  [JC2013a]   –  Refinements:  Flip  interval  dependence  on  value   (binary,  arithme&c)  types,  logis&c  S-­‐curve   discussion  [JC2013b]   22   Graphs  from  [JS2013a]  
  23. “workload” (i.e. program) PMC Sugges&ons  for  Bochspwn   23  

    Par(al  Bayes   model  of   RAM  paging   behavior   (1995)   Theory:  Inves&gate  mul&-­‐core  control   systems  (system  scheduler,  Paging)   and  ‘bring  out’  assump&ons   12  high  level   computa(on   language   paRerns  mined   from  seven   general   applica(on   areas  (green   &blue  rare)   [Asan2009]   Prac&cal:  Find  low  level  assembly   paaern  transla&on  and  inves&gate   suscep&bility  to  double  fetch  and   resul&ng  ‘distor&ons’/  error  
  24. Epilogue:  Methods  ‘Blueprint’   •  Signal  Selec&on   –  Construct

     model  of  system   –  Iden&fy  side  channel  observables    (OS/MA  events  &  others)   –  Scope  SCO  ’s  MINE  proper&es   –  Use  MIC/MINE  sta&s&cs  [Reshef2011]   •  Signal  Representa&on  &  Analysis   –  Octave  (free  but  not  powerful  enough),  MATLAB  (best  choice)   •  Boxplots,  Probability  Plots     •  Toolboxes:  Sta&s&cs,  DSP,  System  Iden&fica&on   –  Machine  Learning   •  Internalize  [Dom2012]   •  Select  ML  procedures  from  [Murph2012]  appropriate  for  and  educed  from  a  system  model  and   the  signals’  macro-­‐proper&es   •  [PMTK3]  for  Bayesian  reasoning/modeling   •  Scien&fic  Experiments  [Mont2012]     •  Specialized  tools:   –  Signal  Analysis:  Eureka  [Lipson2009]   –  Signal  Representa&on:  Viewpoints  [Gazis2010]   24  
  25. Pro-­‐Tip  ML   25   •  Know  what  features  your

     favourite  ML  algo  selects   and  weighs     •  Many  blind  spots  possible   •  Study  Domingos  (2013)!  
  26. Current/upcoming  R  &  D  topics   •  Added  to  the

     talk  is  a  short  WP    with  a   selec&on  of  R  &  D  issues     – Concurrency  Aaacks   – Composi&onal  Security   – Systemic  Computer  Security   •  All  these  in  my  humble  opinion  benefit  from   SCA   26  
  27. Concurrency  Aaacks   •  Even  though  we  increasingly  rely  on

     concurrent  execu&on,  such   programs  are  much  more  difficult  to  write,  test,  debug.     –  Poten&al  for  serious  concurrency  errors  in  many  widespread   concurrent  programs,  enabling  feasible  concurrency  aRacks   •  Many  ‘sequen&al’  defense  techniques  ,  if  unaware  of  concurrent   programming,  are  ineffec&ve   •  Careful  study  of  Bowspwn  and  ROPe  will  yield  insights   27   Findings Implications A majority (24 out of 46) of the concurrency attacks corrupt pointer data. Existing memory safety tools, once made aware of concur- rency, may be able to prevent concurrency attacks that cor- rupt pointer data. 9 concurrency attacks directly corrupt scalar data, such as user identifiers, without compromising memory safety. Few existing defenses handle attacks that directly corrupt scalar data. Many existing defenses become unsafe in the face of concur- rency errors These defenses must consider concurrent execution. The exploitability of a concurrency error highly depends on the duration of its vulnerable window (i.e., the timing win- dow within which the concurrency error may occur). New defense techniques may reduce the exploitability of concurrency errors by reducing the duration of the vulner- able window. Table 1: Summary of Findings.
  28. Composi&onal  Security   28  

  29. Systemic  Computer  Security   •  Mo&va&on:  Flash  Crash  (May  

    2010)  –  see  my  IEEE  SP  ar&cle   •  Automated  black-­‐box  algorithmic   trading:  Johnson  (2013)  “Rise  of   the  Machines”   –  phenomenological  ‘signatures’  of   interac&ng  autonomous  computer   agents  in  real-­‐world  dynamic   (trading)  system   –  All-­‐machine  -me  regime   characterized  by  frequent  `black   swan’  events  with  ultrafast   dura&ons  (<650ms  for  crashes,   <950ms  for  spikes   •  Aggregate    behavior  of  simple   agents  is  unpredictable  in   principle;  no  useful  security   guarantees  anent  dynamics   possible   29   HFT  Nanex  2010  
  30. Systemic  Computer  Security  II   •  Aggregate  behavior  of  simple

     agents  is   unpredictable;  no  useful  security  guarantees   anent  dynamics  possible  [Joh13]  [Bil14]   •  Analysis  of  (side  channel)  event  signatures  in   phase  space;  design  of  circuit  breakers,   graceful  degrada&on,  rec&fiers   •  Relevance  to  Singapore  ‘Smart  Ci&es’  (see   good/bad  example  Songdo,  Portland)   30  
  31. Thank  you   How  Scien&sts  Relax   •  Infrared  spectroscopy

     on  a   vexing  problem  of  our  &mes:   Truly  comparing  apples  and   oranges     31   A  spectrographic  analysis  of  ground,   desiccated  samples  of  a  Granny  Smith   apple  and  a  Sunkist  navel  orange.  Picture   from  [San95]   Thank  you  for  your  -me  and   the  considera&on  of  ideas.     I  appreciate  being  at  SyScan   and  to  finally  visit  Singapore     J    
  32. References  I   [Asan2009]  K.  Asanovic  et  al  “A  view

     of  the  parallel  compu&ng  landscape”,  CACM   52:10,  Oct2009,  pp.  56-­‐67  hap://dl.acm.org/cita&on.cfm?id=1562764.1562783       [Boriah2008]  S.  Boriah  et  al,  “Similarity  Measures  for  Categorical  Data”,  SIAM  red,   30:2,  2008  hap://www-­‐users.cs.umn.edu/~sboriah/PDFs/BoriahBCK2008.pdf     [Goldw2012]    S.  Goldwasser  and  G.  Rothblum,  "How  to  Compute  in  the  Presence  of   Leakage,"  FOCS,  Oct.  2012,  pp.31-­‐40   hap://eccc.hpi-­‐web.de/report/2012/010/download/         [Hund2013]  R.Hund  et  al,  “Prac&cal  Timing  Side  Channel  Aaacks  Against  Kernel  Space   ASLR”  IEEE  S  &  P,  2013  ,  pp.  191-­‐205 hap://www.ieee-­‐security.org/TC/SP2013/papers/4977a191.pdf       [Yang2012]  J.  Yang  et  al,  “Concurrency  aaacks”,  USENIX  HotPar,  2012   haps://www.usenix.org/system/files/conference/hotpar12/hotpar12-­‐final44.pdf       [JC2013a]  M.  Jurczyk  &G.  Coldwind,  “Iden&fying  and  Exploi&ng  Windows  Kernel  Race   Condi&ons  via  Memory  Access  Paaerns”,  SyScan,  April  2013   hap://j00ru.vexillium.org/?p=1695       32  
  33. References  II   [Dom2012]  P.  Domingos,  “A  Few  Useful  Things

     to  Know  about  Machine  Learning”,  CACM  55:10,   Oct  2012,  pp.  78-­‐87  haps://t.co/NsAnRUrPtq       [Demme2013]  J.  Demme  et  al,  “On  the  Feasibility  of  Online  Malware  Detec&on  with  Performance   Counters”  ISCA,  2013,  pp.  559-­‐570  hap://www.cs.columbia.edu/~jdd/papers/isca13_malware.pdf   [Mont2012]  D.  Montgomery,  Design  and  Analysis  of  Experiments,  Wiley  Press,  2012,  ch.  1   hap://higheredbcs.wiley.com/legacy/college/montgomery/1118146921/supp_material/ch01.doc   [Murph2012]  K.  Murphy,  Machine  Learning,  MIT,  2012  hap://www.cs.ubc.ca/~murphyk/MLbook/   [Reshef2011]  D.  Reshef  et  al.  "Detec&ng  Novel  Associa&ons  in  Large  Data  Sets"  Science  334.6062   (2011):  1518-­‐1524   hap://www.sciencemag.org/cgi/rapidpdf/334/6062/1518? ijkey=cRCIlh2G7AjiA&keytype=ref&siteid=sci         [Reshef2011sup]  D.  Reshef  et  al.  SOM  [Reshef2011],  Science  334.6062  (2011) hap://www.sciencemag.org/content/334/6062/1502.full?ijkey=l9Qe0i/ BE6ZOI&keytype=ref&siteid=sci       [Wich2013]  G.  Wicherski  "Taming  the  ROPe  on  Sandy  Bridge”,  SyScan,  April  2013 hap://www.syscan.org/index.php/download/get/3c6891f2e90e661ea23224cd8f419262/ SyScan2013_DAY1_SPEAKER05_Georg_WIcherski_Taming_ROP_ON_SANDY_BRIDGE_syscan.zip       33  
  34. References  III   [JC2013b]  M.  Jurczyk  and  G.  Coldwind,  “Kernel

     double-­‐fetch  race  condi&on  exploita&on  on   x86  –  further  thoughts”,  blog,  June  2013,  hap://j00ru.vexillium.org/?p=1880       [Snyder2008]  L.  Snyder,  “The  whole  box  of  tools:  William  Whewell  and  the  logic  of  induc&on”,   Handbook  of  the  History  of  Logic  (Bri(sh  Logic  in  the  Nineteenth  Century),  Ed.  :  D.  Gabbay,  Vol   4,  2008,  pp.163–228     [Lipson2009]  M.  Schmidt  and  H.  Lipson  "Dis&lling  Free-­‐Form  Natural  Laws  from  Experimental   Data,"  Science  324:5923,  2009,  pp.  81  –  85   hap://ccsl.mae.cornell.edu/sites/default/files/Science09_Schmidt.pdf       [Gazis2010]  P.  Gazis  et  al.,  “Viewpoints:  A  High-­‐Performance  High-­‐Dimensional  Exploratory   Data  Analysis  Tool”,  Publica&ons  of  the  Astronomical  Society  of  the  Pacific,  122(898),  2010,   pp.  1518-­‐1525,  hap://www.giss.nasa.gov/staff/mway/Gazis_Levit_Way2010.pdf       [PMTK3]  K.  Murphy,  M.  Dunham  et  al,  “probabilis&c  modeling  toolkit  for  Matlab/Octave”, 2011,  haps://github.com/probml/pmtk3       [San1995]  S.  Sandford,  “Apples  and  oranges:  a  comparison”,  Annals  of  Improbable  Research   1:3,  1995  hap://www.improbable.com/airchives/paperair/volume1/v1i3/air-­‐1-­‐3-­‐apples.html       [Alistarh2014]  N.  Shavit  et  al,  “Are  Lock-­‐Free  Concurrent  Algorithms  Prac&cally  Wait-­‐Free?”,   2014,  hap://research.microso`.com/pubs/209106/paper.pdf     34