Michael Grønager [ARCHIVE] on Nostr: 📅 Original date posted:2012-02-01 📝 Original message:Hi Gregory, I played with ...
đź“… Original date posted:2012-02-01
đź“ť Original message:Hi Gregory,
I played with the database sync as well to get further speedups, but in the latest version I could only get about 1% extra from this.
In the Satoshi client there is a bunch of sleeps and mutexes (put in there with great generosity) for making threads run smoother and to avoid deadlock I guess.
The big change in speed came from moving from the threading based setup to the async based setup, so I think these are the culprit. Further, as I also write somewhere in the wiki. I can measure that 50% of the time is consumed in verify, so there is not much left to optimize now. Actually, Berkeley DB does a quite decent job in caching reads so not even a cache should help.
Cheers,
M
On 01/02/2012, at 15:59, Gregory Maxwell wrote:
> On Wed, Feb 1, 2012 at 9:18 AM, Michael Grønager <gronager at ceptacle.com> wrote:
>> The libcoin/bitcoind client downloads the entire block chain 3.5 times faster than the bitcoin/bitcoind client. This is less than 90 minutes on a modern laptop!
>
> Very interesting. Do you know where this speedup came from? It's not
> typical for straight refactors that don't change datastructures and
> the like to see such big speedups.
>
> I see you have commented out code that disables fsync, which was my
> first guess since I get big speedups from doing similar things.
Michael Gronager, PhD
Director, Ceptacle
Jens Juels Gade 33
2100 Copenhagen E
Mobile: +45 31 45 14 01
E-mail: gronager at ceptacle.com
Web: http://www.ceptacle.com/
đź“ť Original message:Hi Gregory,
I played with the database sync as well to get further speedups, but in the latest version I could only get about 1% extra from this.
In the Satoshi client there is a bunch of sleeps and mutexes (put in there with great generosity) for making threads run smoother and to avoid deadlock I guess.
The big change in speed came from moving from the threading based setup to the async based setup, so I think these are the culprit. Further, as I also write somewhere in the wiki. I can measure that 50% of the time is consumed in verify, so there is not much left to optimize now. Actually, Berkeley DB does a quite decent job in caching reads so not even a cache should help.
Cheers,
M
On 01/02/2012, at 15:59, Gregory Maxwell wrote:
> On Wed, Feb 1, 2012 at 9:18 AM, Michael Grønager <gronager at ceptacle.com> wrote:
>> The libcoin/bitcoind client downloads the entire block chain 3.5 times faster than the bitcoin/bitcoind client. This is less than 90 minutes on a modern laptop!
>
> Very interesting. Do you know where this speedup came from? It's not
> typical for straight refactors that don't change datastructures and
> the like to see such big speedups.
>
> I see you have commented out code that disables fsync, which was my
> first guess since I get big speedups from doing similar things.
Michael Gronager, PhD
Director, Ceptacle
Jens Juels Gade 33
2100 Copenhagen E
Mobile: +45 31 45 14 01
E-mail: gronager at ceptacle.com
Web: http://www.ceptacle.com/