Fast CCW Translation for Network I/O
Purpose
Our purpose is to measure the effectiveness of the fast CCW translation extensions that have been added to CP in z/VM 4.1.0. This line item is intended to reduce the CP CPU time consumed translating CCWs for guests that do I/O to real CTCs or to LAN Channel Station (LCS) devices. Our experiments showed that for LCS I/O, the line item reduced CP CPU consumption by 39% and total CPU consumption by 25%. For real CTC I/O, the line item reduced CP CPU consumption by 44% and total CPU consumption by 30%.
Hardware
2064-108, LPAR with 2 dedicated engines, 1 GB real, 2 GB expanded, OSA Express Fast Ethernet (LCS mode), wrap-back ESCON channel. MTU 1500 for all cases.
Software
z/VM 4.1.0 with associated TCP/IP; TurboLinux beta 13 (November 2000). 1
LCS Device Experiment
We installed the specified Linux and let it own the LCS device. We placed a 100 MB data file on a nearby workstation on the IBM intranet and set up the Linux guest so that it would do 5 FTP GETs of this file to /dev/null. We measured the Linux guest for virtual CPU consumed, CP CPU consumed, and I/Os performed during the set of GETs. We performed this experiment four times. We computed the mean and standard deviation for the samples.
Next we used the CP STORE HOST command to "zap out" the fast CCW translation path in HCPVDI. In other words, we disabled z/VM 4.1.0's fast CCW translation line item, so that CCW translation would take the customary (z/VM 3.1.0 and earlier) code path. We performed the experiment four more times and computed the means and standard deviations for the samples.
We then performed twin-tailed t-tests over the two sets of samples to look for significance in the difference of the means.
LCS Device Analysis
Measurements are expressed as N/m/sd, where N is the number of samples, m is the mean of the samples, and sd is the standard deviation of the samples.
Result r is our assessment of repeatability: r=1 indicates that the 95% confidence interval on m is contained within the 5% magnitude interval around m.
In the comparisons of means, result cl is the confidence
level the t-test yielded. In other words, it's the certainty we
have that the means truly are different.
Table 1. Fast CCW Translation Results - LCS
| Quantity | Disabled | Enabled | Change |
|---|---|---|---|
| CP CPU (sec) | 4/9.72/0.0277 r=1 | 4/5.91/0.0274 r=1 | cl=99 delta=-39.21% |
| Total CPU (sec) | 4/15.68/0.0526 r=1 | 4/11.7575/0.0396 r=1 | cl=99 delta=-25.03% |
| I/O (count) | 4/306922/701.7 r=1 | 4/307470/1272 r=1 | cl=0 |
| CP CPU per I/O (sec) (calculated) | 3.167E-05 | 1.922E-05 | -39.4% |
| Note: 2064-108, LPAR with 2 dedicated processors, 1 GB main, 2 GB expanded. z/VM 4.1.0. Guests 128 MB; samples / mean / standard deviation | |||
What we see here is that CP CPU time went down 39%, overall CPU time went down 25%, and CP CPU time per I/O went down 39%. There was no change in the number of I/Os started by Linux (we expected that).
Real CTC Experiment
We installed said Linux and connected it via real ESCON (wrap-back) to a VM TCP/IP stack which in turn owned the LCS device. We then ran the same two FTP experiments described previously.
Real CTC Analysis
N/m/sd, r, and cl have the
same meanings as in experiment 1.
Table 2. Fast CCW Translation Results - ESCON
| Quantity | Disabled | Enabled | Change |
|---|---|---|---|
| CP CPU (sec) | 4/14.64/0.1901 r=1 | 4/8.0725/0.0512 r=1 | cl=99 delta=-44.86% |
| Total CPU (sec) | 4/23.5125/0.5512 r=1 | 4/16.4425/0.0810 r=1 | cl=99 delta=-30.07% |
| I/O (count) | 4/325479/3711 r=1 | 4/314870/1086 r=1 | cl=99 delta=-3.26% |
| CP CPU per I/O (sec) (calculated) | 4.498E-05 | 2.564E-05 | -43.0% |
| Note: 2064-108, LPAR with 2 dedicated processors, 1 GB main, 2 GB expanded. z/VM 4.1.0. Guests 128 MB.; samples / mean / standard deviation | |||
We see here that CP CPU time dropped 45%, total CPU time dropped 30%, and Linux I/Os dropped 3%. CP CPU time per I/O dropped by 43%.
Footnotes:
- 1
- uname -a yields Linux linccw1.endicott.ibm.com 2.2.16 #1 SMP Tue Nov 21 12:11:14 PST 2000 s390 unknown.