This is *really* cool Amir, and a good presentation of the data.

Writes are presumably unimportant perforance-wise, because of the
relative cycle times, but have you run read/write tests under load to
validate your concurrency implementation?

My other concern at this point would be fault-tolerance. Have you run
any hard shutdown tests under read and/or write load? I realize block
data is all cache, but obviously we don't want to have to rebuild after
every hard shutdown.

BTW, a side-by-side comparison with bitcoind would be pretty valuable.

e

On 08/27/2014 10:44 AM, Amir Taaki wrote:
> So guys, remember this email?
>
> Well it turns out the performance increase is even more drastic than I
> realised.
>
> https://wiki.unsystem.net/en/index.php/Libbitcoin/Blockchain/htdb_slab_Performance
>
> Read speeds are reduced to ~5% compared to LevelDB with large datasets.
>
> On 27/12/13 00:10, Amir Taaki wrote:
>> I've made significant performance increases for the blockchain by creating
>> a custom database that outperforms the more conventional LevelDB. This is
>> still a work in progress and won't be in libbitcoin at least for some
>> time. You can read more about more research here:
>>
>> https://wiki.unsystem.net/index.php/Libbitcoin/Blockchain
>>
>> The results are drastic. Write speeds are reduced to 30%. Read speeds to
>> 15%. We can even go further as this is an initial prototype. We haven't
>> yet applied all possible optimisations.