Skip to content

Latest commit

 

History

History
145 lines (104 loc) · 7.9 KB

analysis.md

File metadata and controls

145 lines (104 loc) · 7.9 KB
Raw

Bn256Add

In the following order:

  1. cpp_20170824_chfast.txt.md
  2. geth_20170819_mhswende.txt.md
  3. geth_20170821_tim.txt.md
  4. geth_20170822_karalabe.txt.md
  5. geth_20170823_tim.txt.md
  6. geth_20170824_tim_ubuntu.txt.md
  7. parity_20170825_mhswende.txt.md
  8. cpp_20170828_chfast.txt.md
  9. cpp_20170829_chfast.txt.md

This chart shows the worst cases (lowest G/S) for all benchmarks.

Name                                         Gascost         Time (ns)      MGas/S    Gasprice for 10MGas/S    Gasprice for ECDSA eq
-------------------------------------------------------------------------------------------------------------------------------------

bn256Add/chfast2                                 500             8511        58.7475                     85.11            362.026
PrecompiledBn256Add/chfast2                      500            76264         6.55617                   762.64            1333.29
PrecompiledBn256Add/chfast1                      500            65233         7.66483                   652.33            1057.71
PrecompiledBn256Add/chfast2                      500            50981         9.80758                   509.81             964.113
PrecompiledBn256Add/chfast2                      500            43206        11.5725                    432.06             898.135
PrecompiledBn256Add/chfast2                      500            65268         7.66072                   652.68             954.629
bn_128_add                                       500            11291        44.2831                    112.91             236.758
bn256Add/cdetrio13                               500            11879        42.0911                    118.79             404.46
bn256Add/chfast1                                 500             8527        58.6373                     85.27             359.148

For ecdsa-equivalence, a gas cost of 1000 for bn256Add seems appropriate.

Modexp

Order of columns is

  1. geth_20170819_mhswende.txt.md
  2. geth_20170821_tim.txt.md
  3. geth_20170822_karalabe.txt.md
  4. geth_20170823_tim.txt.md
  5. geth_20170824_tim_ubuntu.txt.md
  6. parity_20170825_mhswende.txt.md
  7. cpp_20170828_chfast.txt.md
  8. cpp_20170829_chfast.txt.md
Name                                         Gascost         Time (ns)      MGas/S    Gasprice for 10MGas/S    Gasprice for ECDSA eq
-----------------------------------------  ---------  ----------------  ----------  -----------------------  -----------------------

PrecompiledModExp/nagydani-1-qube                 40    3863              10.3546                     38.63              67.535
PrecompiledModExp/nagydani-1-qube                 40    3844              10.4058                     38.44              62.3277
PrecompiledModExp/nagydani-1-qube                 40    3242              12.3381                     32.42              61.3102
PrecompiledModExp/nagydani-1-qube                 40    2997              13.3467                     29.97              62.2995
PrecompiledModExp/nagydani-1-qube                 40    4393               9.10539                    43.93              64.2533
modexp_1_qube                                     40   12746               3.13824                   127.46             267.268
modexp/nagydani-1-qube                            40   18555               2.15575                   185.55             631.767
modexp/nagydani-1-qube                            40   13026               3.07078                   130.26             548.64

For ecdsa-equivalence, the gasCost for modExp should be adjusted by 1.5, so nagydani-1-qube costs ~60. Note, though, that modExp seems fairly variable, and ranged in MGas/S from e.g. 9 - 66, depending on input vector.

Note : Parity does exponentiation by squaring currently, but plan to optimize it Note2 : CPP seems extremely slow on nagydany-1-qube.

Scalar multiplication

In the following order:

  1. cpp_20170824_chfast.txt.md
  2. geth_20170819_mhswende.txt.md
  3. geth_20170821_tim.txt.md
  4. geth_20170822_karalabe.txt.md
  5. geth_20170823_tim.txt.md
  6. geth_20170824_tim_ubuntu.txt.md
  7. cpp_20170828_chfast.txt.md
  8. cpp_20170829_chfast.txt.md
Name                                         Gascost         Time (ns)       MGas/S    Gasprice for 10MGas/S    Gasprice for ECDSA eq
-----------------------------------------  ---------  ----------------  -----------  -----------------------  -----------------------

bn256ScalarMul/chfast2                          2000   455600              4.38982                  4556              19379.5
PrecompiledBn256ScalarMul/chfast3               2000       2.78246e+06     0.718788                27824.6            48644.5
PrecompiledBn256ScalarMul/chfast3               2000       1.83966e+06     1.08715                 18396.7            29828.9
PrecompiledBn256ScalarMul/chfast2               2000       1.97919e+06     1.01051                 19791.9            37429
PrecompiledBn256ScalarMul/chfast2               2000       1.63471e+06     1.22346                 16347.1            33981.1
PrecompiledBn256ScalarMul/chfast2               2000       2.52246e+06     0.792876                25224.6            36894.3
bn_128_mul                                      2000   683165              2.92755                  6831.65           14325.1
bn256ScalarMul/cdetrio6                          2000  925953              2.15994                  9259.53           31527.2
bn256ScalarMul/cdetrio1                          2000  701889              2.84945                  7018.89           29562.8

For ecdsa equivalence, the gasCost for scalarMul should be ~30K. However, some the benchmarks above did not use worst-case scenarios (cdetrio6 / cdetrio1) , and some new benchmarks show that ~60% longer times are to be expected, which suggests that perhaps 50K would be more appropriate.

The 30K suggestions by Cpp are using the worst-case inputs, whereas the 30K suggestions by geth are not.

Pairing

In the following order:

  1. cpp_20170824_chfast.txt.md
  2. geth_20170819_mhswende.txt.md
  3. geth_20170821_tim.txt.md
  4. geth_20170822_karalabe.txt.md
  5. geth_20170823_tim.txt.md
  6. geth_20170824_tim_ubuntu.txt.md
  7. cpp_20170828_chfast.txt.md
  8. cpp_20170829_chfast.txt.md
Name                                         Gascost         Time (ns)      MGas/S    Gasprice for 10MGas/S    Gasprice for ECDSA eq
-----------------------------------------  ---------  ----------------  ----------  -----------------------  -----------------------

bn256Pairing/ten_point_match_2                900000       4.57002e+07    19.6936                 457002            1.94392e+06
PrecompiledBn256Pairing/one_point             180000       1.84698e+07     9.74564                184698             322899
PrecompiledBn256Pairing/one_point             180000       1.40256e+07    12.8337                 140256             227415
PrecompiledBn256Pairing/ten_point_match_3     260000       2.10154e+07    12.3719                 210154             397426
PrecompiledBn256Pairing/one_point             180000       1.19181e+07    15.103                  119181             247745
PrecompiledBn256Pairing/one_point             180000       1.80948e+07     9.94763                180948             264659
bn_128_pairing                                     1       7.50048e+06     0.000133325            75004.8            157276
bn256Pairing/ten_point_match_2                900000       6.50312e+07    13.8395                 650312            2.21421e+06
bn256Pairing/ten_point_match_2                900000       4.47169e+07    20.1266                 447169            1.88342e+06

For ecdsa equivalence, the pairing gas cost seems like it should be doubled.

The parity-value has a gascost of 1 - I'm not sure what the gascost for that input vector is.

Cpp-results also points towards doubling (or more), in the worst case 900K should be 2.2M.