Michael Bowe wrote:
Hi, I need some hardware/software suggestions for TPROXY servers.
We're an ISP and have been trialling Squid for a while on an assorted
collection of spare hardware. (We've got some Dell 2950 running VMware and
iSCSI, also Dell 2970 with VMware and SAS HDD, also HP DL360-G5 non-VMware
with SAS HDD). All these servers have been working pretty well, but we now
need to work out a budget to buy proper dedicated gear.
We have some POPs which have about 5000 cable modems. Cisco routers running
WCCP feed groups of Squid 3.1 / TPROXY servers. At the moment we have 3 to 4
sibling no-cache servers at each POP.
We currently deal with HTTP traffic of about 150Mbps but would like to
dimension the new gear to support double this.
What are peoples opinions on what sort of hardware to use?
Maybe groups of mid-size servers be best? I was thinking along the lines of
:
Eg HP DL360/380-G6, 1 x 2.4Ghz quad core, 32Gb RAM, 4 x 15K 146Gb HDD
Or should we be looking at just using one larger server :
Eg HP DL380-G6, 1 x 2.4Ghz (or faster) quad core, 64Gb RAM, 8 x 15K 146Gb
HDD
CPU-wise Squid only makes use of one CPU, but having dual-core or more
lets you allocate one CPU to the OS and other bits then to also run one
instance of Squid per extra core. Pinning cores to squid instances is a
plus where possible.
If you get 4+ core hardware, a CARP model with one instance receiving
all requests and balancing across the other cores for actual storage
handling with 1+ disk per core scales extremely well.
I'd be interested in hearing what req/sec you manage to achieve. We only
have reported performances for up to 3.0 so far.
When buying HDDs for these servers, you can choose between 2.5" and 3.5"
drives. Should there be much difference in performance between the two? I
see the 3.5" version of 15K's are quite a bit cheaper and also are available
in sizes > 146Gb
I would not expect a lot. The only thing that really matters for modern
disks is IO speed.
There are choices for the disk controller. Eg HP lets you choose between
256M, 512M, 1G RAM on the supplied P410i RAID card. We wouldn't be running
any RAID, but would extra RAM on the card still be helpful with speeding up
disk access for squid?
If the drives use it. Squid is not very helpful for things like that yet.
Dell R610/R710 seems pretty similar to the HP DL360/380. But with the Dell
you have to buy dual quad core, which seems a bit wasteful for squid?
(unless you were going try and dice the server up to run multiple squids
under VMware). Probably best to save a few $$ and stick with 1 x CPU and put
extra cash towards more RAM or disk? Or do you think VMware is an OK way to
make use of all the CPU's? On our trial servers the VMware ESXi seems to
work OK but in the back of my mind I worry about the extra overhead it
introduces.
More RAM the better. Memory caching is not used to the best efficiency
in Squid but still reaches a few orders of magnitude more requests per
second than disk. Having enough RAM to prevent swapping is an absolute
MUST.
For disks, the faster IO the better and a few smaller disks are better
than one single large disk.
Amos
--
Please be using
Current Stable Squid 2.7.STABLE7 or 3.0.STABLE20
Current Beta Squid 3.1.0.15
