| 1 |
Emulating Netbench |
| 2 |
================== |
| 3 |
|
| 4 |
Netbench is a terrible benchmark, but it's an "industry standard" and |
| 5 |
it's what is used in the press to rate windows fileservers like Samba |
| 6 |
and WindowsNT. |
| 7 |
|
| 8 |
The big problem with netbench for the open source community is that |
| 9 |
very few people who work on open source software have the facilities |
| 10 |
to run it properly. You need a lab with between 60 and 150 windows PC, |
| 11 |
all on switched fast ethernet and a really grunty server (say a quad |
| 12 |
xeon with 1GB ram and hardware raid). Then you need some way to nurse |
| 13 |
all those machines along so they will run a very fussy benchmark suite |
| 14 |
without crashing. Not easy and very expensive. Only one person in the |
| 15 |
open source community that I know of has access to such a lab (Jeremy |
| 16 |
Allison at SGI). |
| 17 |
|
| 18 |
In order for the development methodologies of the open source |
| 19 |
community to work we need to be able to run this benchmark in an |
| 20 |
environment that a bunch of us have access to. We need the source to |
| 21 |
the benchmark so we can see what it does. We need to be able to split |
| 22 |
it into pieces to look for individual bottlenecks. In short, we need |
| 23 |
to open up netbench to the masses. |
| 24 |
|
| 25 |
To do this I have written three tools, dbench, tbench and |
| 26 |
smbtorture. All three read a load description file called client.txt |
| 27 |
that was derived from a network sniffer dump of a real netbench |
| 28 |
run. client.txt is about 4MB and describes the 90 thousand operations |
| 29 |
that a netbench client does in a typical netbench run. They parse |
| 30 |
client.txt and use it to produce the same load without having to buy a |
| 31 |
huge lab. They can simulate any number of simultaneous clients. |
| 32 |
|
| 33 |
client.txt must either be in the working directory, or specified on |
| 34 |
the command line with the -c option. |
| 35 |
|
| 36 |
dbench |
| 37 |
------ |
| 38 |
|
| 39 |
dbench produces only the filesystem load. It does all the same IO |
| 40 |
calls that the smbd server in Samba would produce when confronted with |
| 41 |
a netbench run. It does no networking calls. |
| 42 |
|
| 43 |
You can get dbench from ftp://samba.org/pub/tridge/dbench/ |
| 44 |
You run it as "dbench N" where N is the number of clients to |
| 45 |
simulate. It gives out three numbers like this (this is from a 144 |
| 46 |
client run on a quad xeon box): |
| 47 |
|
| 48 |
Throughput 40.6701 MB/sec (NB=50.8376 MB/sec 406.701 MBit/sec) |
| 49 |
|
| 50 |
the first is the true throughput as seen by dbench. The second and |
| 51 |
third numbers are "netbench scaled" numbers that give the throughput |
| 52 |
that would be seen by Win9X boxes after taking into account the client |
| 53 |
file cacheing performed by oplocks. They are given in both MB/sec and |
| 54 |
MBit/sec as different netbench reports use different scales. |
| 55 |
|
| 56 |
tbench |
| 57 |
------ |
| 58 |
|
| 59 |
tbench produces only the TCP and process load. It does the same socket |
| 60 |
calls that smbd would do under a netbench load. It does no filesystem |
| 61 |
calls. The idea behind tbench is to eliminate smbd from the netbench |
| 62 |
test, as though the smbd code could be made infinately fast. The |
| 63 |
throughput results of tbench tell us how fast a netbench run could go |
| 64 |
if we eliminated all filesystem IO and SMB packet processing. tbench |
| 65 |
is built as part of the dbench package. |
| 66 |
|
| 67 |
To run tbench first run tbench_srv on the server. Then run "tbench N SERVER" |
| 68 |
on the client. N is the number of clients to simulate and SERVER is |
| 69 |
the hostname of the server. The results are given in the same format |
| 70 |
as dbench. |
| 71 |
|
| 72 |
smbtorture |
| 73 |
---------- |
| 74 |
|
| 75 |
smbtorture is the stress tester from the Samba suite. I've recently |
| 76 |
added a stress test that allows you to simulate the full netbench |
| 77 |
benchmark, including all network traffic and filesystem IO. |
| 78 |
|
| 79 |
To run smbtorture you first need to install Samba version 2.0.X. The |
| 80 |
binary distrubtions at ftp://ftp.samba.org/pub/samba/bin-pkgs/redhat/ |
| 81 |
would do fine if you don't want to compile it yourself. Then setup a |
| 82 |
netbench share on the fastest disk you have, making sure you have at |
| 83 |
least 25MB free per simulated client. The simplest smb.conf would look |
| 84 |
something like this: |
| 85 |
|
| 86 |
[global] |
| 87 |
security = share |
| 88 |
|
| 89 |
[netbench] |
| 90 |
path = /data/netbench |
| 91 |
read only = no |
| 92 |
guest ok = yes |
| 93 |
|
| 94 |
Then you need smbtorture. You can either grab a precompiled i386 |
| 95 |
Redhat5.2 binary from ftp://samba.org/pub/tridge/dbench/smbtorture.gz |
| 96 |
or you can follow the instructions at http://samba.org/cvs.html to |
| 97 |
download the SAMBA_2_0 branch of the Samba cvs tree and use |
| 98 |
"make smbtorture" to build it. |
| 99 |
|
| 100 |
Finally, you'll need client.txt from |
| 101 |
ftp://samba.org/pub/tridge/dbench/dbench.tgz in the same directory |
| 102 |
that you run smbtorture from. |
| 103 |
|
| 104 |
To run it you do this: |
| 105 |
|
| 106 |
smbtorture //localhost/netbench -U% -N 32 NBW95 |
| 107 |
|
| 108 |
that will run a 32 client load. You can, of course, also run |
| 109 |
smbtorture against a dfferent SMB server (such as a NT server) to give |
| 110 |
comparitive results for any client load that the server can handle. |
| 111 |
|
| 112 |
Even better is to run smbtorture on one machine and smbd on another, |
| 113 |
connected by a very fast network (such as gigabit ethernet). That will |
| 114 |
stop the smbtorture code itself from loading the server and will also |
| 115 |
test the network driver instead of the loopback driver. |
| 116 |
|
| 117 |
results |
| 118 |
------- |
| 119 |
|
| 120 |
To give you an idea of what to expect here are some results on a quad |
| 121 |
xeon machine running Linux 2.2.9 with 1GB of memory (it has 4GB but |
| 122 |
Linux only uses 1GB by default). The machine also has a 45GB hardware |
| 123 |
raid system and a Alteon AceNIC gigabit network card. |
| 124 |
|
| 125 |
The results below are in netbench MB/sec (the NB= number in the |
| 126 |
result). Multiply by 8 to get the MBit/sec numbers that Mindcraft used |
| 127 |
in their graphs. The first column is the number of simulated clients, |
| 128 |
the second is the result. |
| 129 |
|
| 130 |
With dbench I get: |
| 131 |
|
| 132 |
16 118 |
| 133 |
32 103 |
| 134 |
48 81 |
| 135 |
64 74 |
| 136 |
80 67 |
| 137 |
96 57 |
| 138 |
112 57 |
| 139 |
128 54 |
| 140 |
|
| 141 |
With tbench on loopback I get: |
| 142 |
|
| 143 |
16 31 |
| 144 |
32 30 |
| 145 |
48 29 |
| 146 |
64 29 |
| 147 |
80 28 |
| 148 |
|
| 149 |
With tbench running across the gigabit network (using a dual |
| 150 |
processor Origin200 as the client) I get: |
| 151 |
|
| 152 |
16 29 |
| 153 |
32 29 |
| 154 |
48 29 |
| 155 |
64 29 |
| 156 |
|
| 157 |
|
| 158 |
With smbtorture running over loopback I get: |
| 159 |
|
| 160 |
16 18 |
| 161 |
32 18 |
| 162 |
48 18 |
| 163 |
64 17 |
| 164 |
80 16 |
| 165 |
|
| 166 |
With smbtorture running across the gigabit network (using a dual |
| 167 |
processor Origin200 as the client) I get: |
| 168 |
|
| 169 |
16 16 |
| 170 |
32 20 |
| 171 |
48 21 |
| 172 |
64 20 |
| 173 |
80 19 |
| 174 |
|
| 175 |
With smbtorture running across the gigabit network but with smbd |
| 176 |
modified so that write_file() and read_file() are null operations |
| 177 |
(which eliminates the file IO) I get: |
| 178 |
|
| 179 |
16 17 |
| 180 |
32 22 |
| 181 |
48 24 |
| 182 |
64 24 |
| 183 |
80 24 |
| 184 |
|
| 185 |
|
| 186 |
The above results show that, at least for this hardware configuration, |
| 187 |
the problem isn't the filesystem code or the raid drivers. More tests |
| 188 |
will be needed to find out exactly what the problem is but it looks |
| 189 |
like a TCP scaling problem. |
| 190 |
|
| 191 |
comparison with NT |
| 192 |
------------------ |
| 193 |
|
| 194 |
Hopefully Jeremy will be able to run smbtorture against NT on the same |
| 195 |
hardwre sometime in the next week so we have direct numbers for |
| 196 |
comparison, but from looking at the mindcraft numbers under netbench |
| 197 |
we would expect NT to get about the following: |
| 198 |
|
| 199 |
16 8 |
| 200 |
32 14 |
| 201 |
48 21 |
| 202 |
64 27 |
| 203 |
80 31 |
| 204 |
|
| 205 |
so we do well by comparison with small client loads but fall behind |
| 206 |
quite a lot with large loads. Note that the numbers in the mindcraft |
| 207 |
report for Linux/Samba are quite a long way behind what I give above |
| 208 |
because mindcraft did a hopeless job of tuning Linux/Samba. |
| 209 |
|
| 210 |
comparison with netbench |
| 211 |
------------------------ |
| 212 |
|
| 213 |
An initial comparison with real netbench results shows that smbtorture |
| 214 |
does produce very similar throughput numbers. They aren't exactly the |
| 215 |
same but they are similar enough for us to target our tuning efforts |
| 216 |
and expect to see improvements reflected in real netbench runs. When |
| 217 |
we find something that looks promising we can get Jeremy to run a real |
| 218 |
netbench test. |
| 219 |
|
| 220 |
|
| 221 |
a quick hack |
| 222 |
------------ |
| 223 |
|
| 224 |
the tbench results really pointed at the problem being the Linux TCP |
| 225 |
stack. I made a quick (and very unsafe!) hack to Samba and the Linux |
| 226 |
kernel to see if I could remove the lock_kernel() in sys_sendto() and |
| 227 |
sys_recvfrom() for smbd processes by passing a MSG_NOLOCK flag in |
| 228 |
send() and recv(). That gave an enormous improvement in the loopback |
| 229 |
tbench results: |
| 230 |
|
| 231 |
16 31->51 |
| 232 |
32 30->49 |
| 233 |
48 29->47 |
| 234 |
64 29->46 |
| 235 |
80 28->44 |
| 236 |
|
| 237 |
and in the loopback smbtorture results I also saw a big improvement: |
| 238 |
|
| 239 |
16 18->26 |
| 240 |
32 18->27 |
| 241 |
48 18->26 |
| 242 |
64 17->25 |
| 243 |
80 16->24 |
| 244 |
|
| 245 |
that's a 50% improvement. I suspect the numbers will be higher with a |
| 246 |
real netbench run as it won't have the overhead of running 80 |
| 247 |
smbtorture clients on the same box as the server. |
| 248 |
|
| 249 |
|
| 250 |
realistic load? |
| 251 |
--------------- |
| 252 |
|
| 253 |
One question some people may ask is whether the above represents a |
| 254 |
realistic load on a fileserver. It doesn't. Nearly 90% of the |
| 255 |
read/write IO operations in netbench are writes whereas in a "normal" |
| 256 |
office load reads dominate. Also, the load is *much* higher than a |
| 257 |
normal office PC would put on a server. There aren't many office PCs |
| 258 |
that write over a hundred megabytes of data to a server in a few |
| 259 |
minutes, unless maybe they are copying a CD. |
| 260 |
|
| 261 |
That doesn't mean the benchmark is useless, it just means you |
| 262 |
shouldn't use this for purchasing decisions unless you really |
| 263 |
understand the results and how they relate to your environment. |
| 264 |
|
| 265 |
|
| 266 |
license |
| 267 |
------- |
| 268 |
|
| 269 |
smbtorture and dbench are released under the terms of the GNU Public |
| 270 |
License version 3 or later. |
| 271 |
|
| 272 |
|
| 273 |
Andrew Tridgell |
| 274 |
[email protected] |