1. 19 Oct, 2017 1 commit
  2. 17 Oct, 2017 4 commits
  3. 16 Oct, 2017 8 commits
  4. 15 Oct, 2017 2 commits
  5. 13 Oct, 2017 13 commits
  6. 12 Oct, 2017 4 commits
    • Test's avatar
      X neo1 - neo2 timings; delta tgroX, txc · 9cab8f1f
      Test authored
      compared to tgro5:
      
      tgro0   py -> noise (ZEO, NEO/pylite, NEO/pysql, partly neo/go)
      tgro10  py=no-noise, tpy~=tgro5; tgo serial~same, +prefetch more noise (?) 40|65
      tgro15  lat_tcp 1B 49->45 sometime (consistently) otherwise more-or-less same
      tgro20  lat_tcp 4K 170µs->167µs (c)
      tgro25  lat_tcp 4K 170µs->167µs (c), lat_tcp 4K 175 -> 171µs (go)
              ZEOv a bit
              go-go + prefetch !sha1 -> noise + higher (~65µs -> 70 - 110µs)
      
      txc40µs4f seems to be worse than txc200µs4f
      9cab8f1f
    • Test's avatar
      X time neo1 - neo2; δ: gro_flush_timeout=5µs, txc=20μs/0f/0μs-irq/0f-irq · 16f249f1
      Test authored
      Link and TCP* latencies are ~ same and stable.
      
      Compared to b288e62f py timings stabilize:
      
      ZEO		~580-1045μs	-> 585µs
      NEO/pylite	~600-700μs	-> ~550-570µs	(Cpy)
      NEO/pylite	~525-580μs	-> ~450µs	(Cgo)
      NEO/pysql	~820-930μs	-> ~840µs	(Cpy)
      NEO/pysql	~740-800μs	-> ~740µs	(Cgo)
      
      Go timings get a bit worse:
      
      NEO/go		~360µs		-> ~380µs	(Cpy)
      NEO/go		~160µs		-> ~165-170µs	(Cgo)
      NEO/go-nosha1	~140µs		-> ~150µs
      
      Compared to 15a9ccef go+pretech128 timings stabilize:
      
      go-go+prefetch128		~65-160µs	-> ~40-45µs, 60µs(x1)
      go-go+prefetch128 (!sha1)	~60-150µs	-> ~60µs,    40µs(x1)
      16f249f1
    • Kirill Smelkov's avatar
      X show gro_flush_timeout + friends · 68d5b015
      Kirill Smelkov authored
      68d5b015
    • Kirill Smelkov's avatar
      . · 86ab45eb
      Kirill Smelkov authored
      86ab45eb
  7. 11 Oct, 2017 1 commit
    • Test's avatar
      X reran neo1 - neo2 after reloading r8169 driver on both machines · 15a9ccef
      Test authored
      Compared to run from Oct10 txc settings were not changed and they are now
      200μs/4f/0μs-irq/0f-irq on both machines instead of 200μs/4f/0μs-irq/0f-irq on
      neo1 and 200μs/0f/0μs-irq/0f-irq on neo2.
      
      The noise about py runs remains.
      
      Go-go serial time remains the same ~155μs without much noise, but Go-go time
      with +prefetch128 gets worse:
      
      NEO/go Cgo +prefetch128
      
              ~45μs   -> 65 - 160μs
      
      Probably it makes sense to tune txc too...
      15a9ccef
  8. 10 Oct, 2017 7 commits
    • Test's avatar
      X neo1 - neo2 timings with turned on TSO, SG, TX · b288e62f
      Test authored
      See previous commit for details about why latency is bad for TCP payload > MSS
      without TSO.
      
      Compared to latest neo1-neo2 timings from Oct05 (C-states disabled, no
      rx-delay) it improves:
      
      TCP1             ~45μs  ->   ~45μs
      TCP1472         ~430μs			# TCP lat. anaomaly
      TCP1400                 ->  ~120μs	# finally
      TCP1500                 ->  ~130μs	# fixed
      TCP4096         ~285μs  ->  ~170μs	# !
      
      ZEO             ~670μs  ->  ~580-1045μs (?)
      NEO/pylite      ~605μs  ->  ~600-700μs  (?)     (Cpy)
      NEO/pylite      ~505μs  ->  ~525-580μs  (?)     (Cgo)
      NEO/pysql       ~900μs  ->  ~820-930μs  (?)     (Cpy)
      NEO/pysql       ~780μs  ->  ~740-800μs  (?)     (Cgo)
      NEO/go          ~430μs  ->  ~360μs              (Cpy)	# <-- NOTE
      NEO/go          ~210μs  ->  ~160μs              (Cgo)	# <-- NOTE
      NEO/go-nosha1   ~190μs  ->  ~140μs			# <-- NOTE
      
      not sure about noise in pure py runs but given raw tcp latency absolutely
      improves this should be a good change to make.
      b288e62f
    • Kirill Smelkov's avatar
      X neotest: Show NIC features and emit warning if !TSO · 4c815af9
      Kirill Smelkov authored
      On neo1 - neo2 without TSO latency becomes very poor in lat_tcp when
      payload size becomes greater TCP MSS (lat_tcp -m 1448 ~ 130μs; lat_tcp
      -m 1449 ~ 500μs and more)
      4c815af9
    • Kirill Smelkov's avatar
      4436b983
    • Kirill Smelkov's avatar
      . · 474c17bc
      Kirill Smelkov authored
      474c17bc
    • Kirill Smelkov's avatar
      X time for z6001 (localhost) and z6001-z600 · b6d6344d
      Kirill Smelkov authored
      Same a before with added information.
      In particular z6001-z600 shows there is no TCP RR step-wise 400μs increase
      while going 1400B -> 1500B as it is currently the case on RTL.
      b6d6344d
    • Kirill Smelkov's avatar
      . · 50b0d130
      Kirill Smelkov authored
      50b0d130
    • Test's avatar
      X neo1 - neo2 timings · c5df0f47
      Test authored
      Same as before but with more information and adjusted ping and TCP RR.
      
      In particular TCP RR shows:
      
      TCPRR1400	~ 120μs
      TCPRR1500	~ 430μs
      
      The time when this increases step-wise is when TCP packet crosses fitting 1 ethernet frame.
      It does not happen on z600* with their current settings.
      c5df0f47