FoundationDB

Cluster tuning cookbook


(Alex Miller) #21

Correct. (In the conf file, and on the command line.)

That’s correct though. The majority of proxy memory usage comes from processing and handling commits, and SERVER_MEM_LIMIT caps that. It is not a generic memory limiting knob though, unlike what its name implies.


(Christophe Chevalier) #22

Setting knob_server_mem_limit = ... in foundationdb.conf does not seem to be supported: the corresponding fdbserver process crashes with Process="fdbserver.4503": Unrecognized knob option 'server_mem_limit'

It seems that this particular knob is supposed to be configured via the --memory command line argument, which corresponds to the memory = ... field in foundationdb.conf. So I think I’m back to square one :slight_smile:


(Alex Miller) #23

It appears the change that adds knob_server_mem_limit isn’t in 5.2, and I forgot that would be a relevant detail for you. :confused:


(Christophe Chevalier) #24

Oh, right. And I was also looking at the code in the master branch :slight_smile:


(Amirouche) #25

IMO the development part with foundationdb is a breeze but the operation part is daunting. I did not read the implementation details of so that’s maybe why I have a difficult time to understand the conversation around spinning a FDB cluster.


(A.J. Beamon) #26

I believe the default is 2GB, which makes the page cache one of the bigger users of memory. There is also a 64k page cache which is 200MB.


#27

Hi! I’ve started playing around with FoundationDB with an aim towards understanding its storage engine performance characteristics, and this thread (and the forum in general) is of great help! I’m using a similar test parameter set to test R/W performance (started with a single FDB instance, and single ssd or single memory engines):
"
testTitle=RandomReadWriteTest
testName=ReadWrite
testDuration=30.0
transactionsPerSecond=1000000
readsPerTransactionA=10
writesPerTransactionA=0
readsPerTransactionB=0
writesPerTransactionB=10
alpha=${rwmix}
nodeCount=10000000
valueBytes=100
"
However, I needed to increase the nodeCount to 10000000 to get ~1GiB “Sum of key value sizes” as reported in the fdbcli status.