The kernel slows execution of code by about factor of 2

[edgarcosta]

Try the code

sieve(n) =
{
    my(S,p); /* necessary to avoid creating global S,p */
    S = apply((x)->1,[1..n]);
    for(p=2,floor(sqrt(n)),
        if(S[p],forstep(m=p+p,n,p,S[m]=0));
    );
    return(select((x)->S[x],[2..n]));
};
c = gettime();
w = getwalltime();
print(sieve(10^6)==select((p)->isprime(p),[1..10^6]));
print("Time: ", strtime(gettime() - c));
print("Time: ", strtime(getwalltime() - w));
c = gettime();
w = getwalltime();
print(length(sieve(10^7)));
print("Time: ", strtime(gettime() - c));
print("Time: ", strtime(getwalltime() - w));

in the kernel and in the terminal.
In the terminal I obtain:

Time: 2,535 ms
Time: 2,644 ms
  *** apply: Warning: increasing stack size to 1024000000.
  *** select: Warning: increasing stack size to 2048000000.
664579
Time: 31,440 ms
Time: 32,261 ms

and in the Kernel

Time: 4,479 ms
Time: 5,030 ms
  ***   Warning: increasing stack size to 1024000000.
  *** select: Warning: increasing stack size to 1200001024.
664579
Time: 51,229 ms
Time: 52,096 ms

any insight on how to fix this?

My alternative is to have a kernel based on pexpect, but then I can't easily support completion

[videlec]

Easier test case: quadclassunit(1 - 2^110) and quadclassunit(1 - 2^120) (take respectively ~3sec and ~15sec).

• I confirm the x2 slowdown compared to GP
• It turns out that execution with cypari2 inside IPython shows the same x2 slowdown... but does not show up in CPython

There is a intriguing difference of behaviour if we load sage or not inside CPython

>>> from cypari2 import Pari
>>> from time import time
>>> pari  =Pari()
>>> pari.allocatemem()
PARI stack size set to 16007168 bytes, maximum size set to 16007168
>>> t0 = time(); s = pari("quadclassunit(1 - 2^110)"); print(time() - t0)
2.941565990447998

Get twice slower if we load sage

$ python
Python 3.10.8 (main, Nov  1 2022, 14:18:21) [GCC 12.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from sage.all import *
>>> from time import time
>>> from cypari2 import Pari
>>> pari = Pari()
>>> t0 = time(); s = pari("quadclassunit(1 - 2^110)"); print(time() - t0)
5.750366926193237

Since the PARIKernel inherits from ipykernel.kernelbase.Kernel it seems plausible that the problem is related to IPython.

[edgarcosta]

I ended up solving this on my end by using pexpect and reading temporary files for each cell https://github.com/edgarcosta/gp_kernel

…

On Thu, Dec 8, 2022, 02:18 Vincent Delecroix ***@***.***> wrote: I confirm the x2 slowdown compared to GP... — Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACO2BSY4DK7IDDZLHZUFETWMGYHNANCNFSM53IDQLPQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>