On Dec 7, 2009, at 7:23 PM, Jonathan Ellis wrote:

> same thing, you are going to need multiple threads to max it out

I created up to 100 threads and read randomly. Some speed up but not actually 
mentionable. The threads didn't load the CPU mentionably either. I noticed that 
the thread dump was full of Thift (TBinaryProtocol something) 
InputStream.read(), 10MiB constant read from HD.

> but yes, reads are typically slower than writes in cassandra because
> of how the log-based merge structures work
> 
> On Mon, Dec 7, 2009 at 11:09 AM, Timo Nentwig <[email protected]> wrote:
>> 
>> On Dec 7, 2009, at 5:59 PM, Jonathan Ellis wrote:
>> 
>>> yes and no -- that's about 4200/s, which is typical for only a single
>> 
>> When writing, yes. But I would expect reading to be much faster (?). 
>> Re-executing the read test doesn't speed up things either (I/O caches).
>> 
>>> thread but 1/3 to 1/5 of what you'd expect it to max out (on our
>>> quad-core test boxes) when you add client threads
>>> 
>>> On Mon, Dec 7, 2009 at 10:38 AM, Timo Nentwig <[email protected]> 
>>> wrote:
>>>> Hi!
>>>> 
>>>> I just downloaded, installed, start cassandra and ran very simple 
>>>> "benchmark": write n times something with 
>>>> key==value==testInsertAndGetAndRemove_n (one thread).
>>>> 
>>>> For n==10 million on a 7200rpm HDD (4G RAM - there should have be 
>>>> "reasonably" free mem however I didn't check) this took 40min (insert()ing 
>>>> one after another). Reading them one by one in sequence delivers about 
>>>> 100/s, reading in 1.000er batches (i.e. multigetColumn()) takes 5-10s 
>>>> (depending on n, the higher the slower).
>>>> 
>>>> Are this typical numbers for cassandra (0.5)? I actually took the 
>>>> configuration as it was.
>> 
>> 

Reply via email to