selecting optimal threshold for svm classificatin

hello frds

usually the output of binary svm is scores ,mostly positive class gets high score whereas low for negative class ,upto this ok , my problem is , if i give a input, the system has to tell whether positive or negative , for that we need a thershold , i dont know how to select this threshold , i tried roc , perfcurve in matlab ,from perfcurve, how can i select best threshold . i confused , ur help can make me easy to solve this. thanks to everybody