Study Notes

In order to resolve UT issue, mounting big volumes is very slowly and the customer set the maximum time no more than 3 seconds. In fact, we find that every mount version on 2.6 kernel will mount the volumes very quickly. However, if the kernel is 2.4, it is expected to have pain. We do not want to modify our code in kernel level, so that have to test the whole system performance when nfsd is running, though usually 4-8 nfsd threads are default. The only reason is single nfsd thread could not cause “dlm_sendd” eat all the CPU resources.

Five servers which implemented with Gigabyte network device, and all the cables are connected to a dedicated Cisco switch in order to avoid network traffic. Every server held 10-12 iozone threads to write data across two NAS 880 nodes. The performance is quite good,single node “bo” is around 25MBps, and during the 3 hours continuous iozone pressure testing the node CPUs usually have high free idles.

It seems the single nfsd would resolve the UT issue now. However, I need confirmation and agreement from Gavin Huang. Thanks Lin Peng and Zhu Feng, you have done great job all these days.

Inventec issue: Their engineers will arrive Shenzhen tomorrow morning to fix SS2280 platforms.