Can a switch give me more speed (long and detailed questions)?
- From: helbig@xxxxxxxxxxxxxxxxxxxxxxxx (Phillip Helbig---remove CLOTHES to reply)
- Date: Sun, 1 Oct 2006 13:53:47 +0000 (UTC)
At present, my cluster is set up like this. I've included ONLY the
information I think is relevant to my question. If I've left something
out, let me know.
o 2 VAXes (VAX 4000/105A and VAXstation 4000/90A) as boot nodes
o 1 ALPHA (DEC 3000/600) as boot node
o 1 ALPHA (ALPHAserver 1200, actually "DIGITAL Server 5000 Model
5305 6533A 5/533 4MB") as satellite
o main user disk is a pair of 9-GB SEAGATE SX910800N disks, one
connected to each of the VAXes
o interconnect for all traffic is the twisted-pair LAN; hub is a
NETGEAR EN108TP, which is an 8-port 10base-T hub)
o also on the HUB is a very modern laptop (not running VMS)
o the uplink from the hub goes to a LINKSYS BEFSR41 which as far
as I know has the LAN ports connected via a SWITCH (also going
into the LinkSys is a VOIP adapter---it can also connect directly
to my ISP by PPPoE, i.e. I could get by without the LinkSys, but
at present it is "behind" the LinkSys just like the hub is)
Since my DSL speed is about to increase from 6000 kb/s to 16,000 kb/s
downstream (present upstream is, I believe, 512 kb/s; I'm not sure what
it will be after the upgrade, but that's not relevant here), I'm
considering to what extent replacing my hub with a switch would speed up
things, whereby there are two distinct questions (both related to the
two things where I would actually notice more speed):
o Could a switch speed up shadow copies? Since the VAX SCSI
speed is 5 Mb/s, if I've done the math right a transfer should
take 4 hours (9*1000*8/5/60/60). Actually, it takes about twice
as long. On the other hand, presumably this is correct since
first the source is read then the copy is written. Thus, the
bottleneck is probably the VAX SCSI speed and, though there are
some collisions during a shadow copy, they are probably negligible
as far as their effect on the speed goes. Thus, I think the
answer here is "no". (The SEAGATE disks themselves I think are
capable of at least 10 MB/s, but the only faster SCSI connections
I have are on ALPHA. If I had two ALPHAs in the cluster, this
might give me some more speed, if I connected the SEAGATEs to
them. On the other hand, thanks to MINICOPY a full copy only
takes place rarely, for example after an unplanned power outage.
Since I want to be able to do a planned reboot of either host
node without a full copy, if I connect one SEAGATE to an ALPHA, I
would have to have another ALPHA in the cluster for MINICOPY to
work.)
o My ISP has a "test your speed" link (which anyone here can try
out): http://www.1und1.de/index.php?page=speedtest . (It
essentially has a lot of garbage as an HTML comment then some code
to check the download speed. I think it is OK as far as reporting
the speed, but the graphical scale is non-linear. Trust the
numbers, not the speed-bar graphics.) The following numbers are
reproducible: MOZILLA on the satellite: 1000-1100 kb/s; LYNX on
the satellite: 3000-4000 kb/s; laptop: 5000-6000 kb/s; either
VAX: 1400-1500 kb/s; LYNX on the 3000/600: 3300--3500 kb/s. The
speed indicated by LYNX agrees with the speed indicated by the
web page. At first, I thought the laptop was getting the full
speed (it probably is) and Mozilla on the satellite was suffering
by having to write the file not to a local disk but to the shadow
set on the VAXes, thus potentially suffering from a) the slow
SCSI speed, b) the fact that it has to write to both disks, c)
the fact that the write is across the network and d) the slow and
as far as I know non-duplex ethernet cards on the VAXes. However,
LYNX on the same machine as Mozilla is quite a bit faster. It
could be that the speedup comes from ignoring the images, though.
OK, this is just about plausible. If the ethernet card in the
3000/600 is better than in the VAXes, then I could see why the
former speed is better than the latter, and if the ethernet card
in the satellite is better than in the 3000/600, that would
explain why the satellite speed (with LYNX) is best of all (except
for the laptop, which has the best ethernet card of all and saves
the file to a local (very fast) disk). Based on these numbers,
how should they change under the following three scenarios: a)
switch instead of hub, b) 16,000 kb/s DSL (this might depend on
whether the WAN speed of the router is limited to, say 10 Mb/s)
and c) both a) and b). The "problem" is the difference between
the laptop, which essentially gets full speed, and Mozilla, which
is only getting about 1/6 the full speed. Obviously, I would like
to get the full speed not only now but also after moving to 16,000
kb/s (assuming the LinkSys is not the bottleneck then) from
Mozilla on the satellite, since this is the one application where
I really would notice more speed on a regular basis. The question
is, of all the things slowing it down (a) to d) above), which, if
any, could be improved by a switch (probably only d) and how much
would that improvement be? Perhaps folks can report the speed
indicated by the web page above, note their cluster (or
single-machine) setup and see if the numbers are plausible.
Note that the first question doesn't have anything to do with the DSL
connection. Rather, I'm wondering whether, if the answer to the second
question is "yes" and I thus buy a switch, a side-effect would be a
speedup as described in the first question.
Bottom line: is it worth it to me to replace the hub in my cluster with
a switch, or are there so many other bottlenecks that I wouldn't notice
it? Note that the laptop is on the HUB and gets (almost) the full
speed.
.
- Follow-Ups:
- Re: Can a switch give me more speed (long and detailed questions)?
- From: Dan Foster
- Re: Can a switch give me more speed (long and detailed questions)?
- From: Phillip Helbig---remove CLOTHES to reply
- Re: Can a switch give me more speed (long and detailed questions)?
- Prev by Date: Re: Cluster-wide logicals: multi-architecture cluster
- Next by Date: Re: Cluster-wide logicals: multi-architecture cluster
- Previous by thread: Re: HP could learn from Island
- Next by thread: Re: Can a switch give me more speed (long and detailed questions)?
- Index(es):
Relevant Pages
|
