Hi Loic, At present, I test different stripe-width(4k,16k,32k,256k,512k, 1024k) with k=3/m=2. I used cosbench+librados to test the performance with 10MB-write(object size). Stripe-width avg-cpu-uitl-of-osd(user+kernel) Throghput 4k 83% 90.01 16k 79% 105.01 32k 76% 106.7 64k 72% 107.69 256k 68% 106.52 512k 68% 103.12 1024K 69% 102.57 >From the result, we can know 4k not the best. I know if the stripe-width increase, we should pad more data. Is there other considerations? At present, we can't set different stripe-width for different erasure-code pool. Why? Thanks! Jianpeng Ma
 |