1 |
W dniu sob, 21.10.2017 o godzinie 10∶01 +0200, użytkownik Paweł Hajdan, |
2 |
Jr. napisał: |
3 |
> On 20/10/2017 18:15, Michał Górny wrote: |
4 |
> > W dniu pią, 20.10.2017 o godzinie 17∶42 +0200, użytkownik Paweł Hajdan, |
5 |
> > Jr. napisał: |
6 |
> > > Curious, do we have any measurements/estimates of the performance cost? |
7 |
> > |
8 |
> > With a single thread serial processing of all hashes, it's just sum of |
9 |
> > times involved in every hash, i.e. Th = T1 + T2 + T3 + ... You'd have to |
10 |
> > get some numbers to get something smarter out of it. |
11 |
> > |
12 |
> > If we assume we can do N threads, then cost of N algorithms is equal to |
13 |
> > the slowest of them all. Which implies that having N algorithms is |
14 |
> > fastest on systems capable of at least N threads. |
15 |
> > |
16 |
> > Taking a random comparison [1], it seems that SHA3/512 is 3-5 times |
17 |
> > slower than SHA2/512. |
18 |
> |
19 |
> How large part of dependency calculation / other portage's operation is |
20 |
> this though? |
21 |
> |
22 |
> My point is, did profiling turn out hash computation as bottleneck, or |
23 |
> is this more speculative? |
24 |
|
25 |
Purely speculative. |
26 |
|
27 |
> I'm still in favor of modernizing the hashes, just somewhat skeptical |
28 |
> when performance is being mentioned. |
29 |
> |
30 |
|
31 |
FWICS BLAKE2 can be even 2.5 times faster than SHA2, so we'll probably |
32 |
go with that. In this case, the performance impact will be negligible -- |
33 |
in fact, it should be faster than the current set of three hashes. |
34 |
|
35 |
-- |
36 |
Best regards, |
37 |
Michał Górny |