|
Figure 36 – A comparison between UNC paths and mapped drives accessing a local Quantum Atlas 10K drive
|
| bet | 44/47 | | Sana | 30.03.2021 | | Hajmi | 1.33 Mb. | | #13789 |
Figure 36 – A comparison between UNC paths and mapped drives accessing a local Quantum Atlas 10K drive. Accessing files using both UNC and mapped drives adds additional system layers and overhead to each request. Compared with single disk results, this results in degraded throughput. Although write performance is the same for both UNC and mapped drives, read performance differs substantially. For good performance, mapped drives should be used.
Accessing local files through UNC and mapped drive paths forces the IO to go through the network stack despite the fact that the IO is local. This combined with some sub-optimal buffering and serialization choices result in poor performance. Since programs will increasingly be using network-transparent names from the Active Directory, this will be an increasingly important issue.
The primary cost of UNC and mapped drives is processor time. Compared with single disk overhead, requests to UNC and mapped drives can cost ten times more processor time per MB. On our machine, 2KB unbuffered mapped reads consumed over 71% of both processors. On small request sizes, UNC showed significantly less overhead for small 2KB writes than on mapped drives. Mapped drives had a fixed overhead of 563μs per write request while UNC paths had a fixed overhead of 454μs per write. 2KB reads showed a smaller impact. As the request size increases, both UNC and mapped drives converge to 27ms per MB.
|
| |
