|
Figure 37 – Unbuffered UNC and mapped overhead
|
| bet | 45/47 | | Sana | 30.03.2021 | | Hajmi | 1,33 Mb. | | #13789 |
Figure 37 – Unbuffered UNC and mapped overhead. UNC and mapped IO are expensive, with up to ten times as much overhead per MB compared with local drives. UNC showed lower overhead than mapped drives for small requests, however this difference in overheads diminishes with increased request size. This implies a higher fixed overhead for mapped drives. We’ve truncated the left graph’s read overhead at the 16KB request size for reads. This is due to the large performance degradation that’s seen with UNC reads with large request sizes. As the number of IOs per second dropped to single digits, the error bound on our soaker became larger than the results, so we were unable to measure that overhead.
CIFS read throughput over Gigabit Ethernet was very good. We were able to achieve over 37.5MBps using sequential 32KB read requests. At 1GBps, the PAP of Gigabit is 125MBps. In round numbers, the half-power point is 63MBps. Although we were unable to achieve the half-power point, this is likely due to the PCI bus as our machines did not reach processor or disk saturation. At 37.5MBps, the server must read from the SCSI controller at 37.5MBps and then write to the Gigabit Ethernet adapter at 37.5MBps. The total bandwidth being used by both reads and writes is 37.5 * 2 or 75MBps, which is nearly the maximum throughput we were able to achieve using local disks.
Unbuffered write throughput was less than half the read throughput. Using 16KB requests, we were only able to achieve 15MBps. With requests larger than 16KB, a performance hit similar to the local UNC case shown in Figure 34 is seen. This indicates that the poor write performance may be due in part to the drive mapping translation. In addition, unlike reads requests that stream data from the server with little acknowledgement from the client, writes require more handshaking as we observed much higher bi-directional network traffic with writes.
Processor utilization for reads is less per MB than that for writes. Whichever system was having data pushed to it had the higher overhead. For reads from the server, processor utilization on the server was 28% compared to 32.2% on the client for 2KB four-deep read requests. For 128KB requests, the server processor utilization was 14% while the client processor utilization was 23.4%. For writes to the server, the server processor utilization was 11.6% for 2KB writes and 7% for 128KB writes, while the client’s processor consumption was 6% and 9.3% respectively.
|
| |
