Hi!
Pretty common task. SCSI initator driver (SCSI miniport or SCSI port) with
TDI lower edge. iSCSI protocol in separate kernel mode DLL. 7-8 MB/sec
(RAW disk read/write) for 100 Mbit cross-over hubless connection. Is this
good or bad? I mean speed. Can anybody tell me what performance boost
we’re expected to get when we’ll migrate from SCSI miniport that use
“tandem SRB completion” to fully deserialzed monolithic SCSI port with TDI
called at DISPATCH level? Does the time invested into development will
cover the resulting performace? [ Will test with 400 Mbit FirWire and GE
soon ]
Any ideas, iSCSI gurus?
Thanks!
Anton
That’s pretty bad. It should be pretty easy to saturate a 100Mb
full-duplex link which would give you ~20MB/sec.
What’s your CPU at? It can’t be very high at that kind of throughput
unless something is terribly wrong. My guess is this is a latency issue.
How many i/o’s do you have outstanding?
----- Original Message -----
From: “Anton Kolomyeytsev”
To: “NT Developers Interest List”
Sent: Thursday, July 25, 2002 9:22 AM
Subject: [ntdev] Performance of iSCSI
> Hi!
>
> Pretty common task. SCSI initator driver (SCSI miniport or SCSI port) with
> TDI lower edge. iSCSI protocol in separate kernel mode DLL. 7-8 MB/sec
> (RAW disk read/write) for 100 Mbit cross-over hubless connection. Is this
> good or bad? I mean speed. Can anybody tell me what performance boost
> we’re expected to get when we’ll migrate from SCSI miniport that use
> “tandem SRB completion” to fully deserialzed monolithic SCSI port with TDI
> called at DISPATCH level? Does the time invested into development will
> cover the resulting performace? [Will test with 400 Mbit FirWire and GE
> soon]
>
> Any ideas, iSCSI gurus?
>
> Thanks!
>
> Anton
>
> —
> You are currently subscribed to ntdev as: xxxxx@alacritech.com
> To unsubscribe send a blank email to %%email.unsub%%
>
> That’s pretty bad. It should be pretty easy to saturate a
100Mb full-duplex link which would give you ~20MB/sec.
Have you ever seen a real initiator/target pair working on such troughput?
You’d need an almost-zero-latency target and good mix of reads and writes in
your test pattern to get this number.
7-8, maybe 10Mb/s sounds more realistic.
What’s your CPU at? It can’t be very high at that kind of
throughput unless something is terribly wrong. My guess is
this is a latency issue. How many i/o’s do you have outstanding?
----- Original Message -----
From: “Anton Kolomyeytsev”
> To: “NT Developers Interest List”
> Sent: Thursday, July 25, 2002 9:22 AM
> Subject: [ntdev] Performance of iSCSI
>
>
> > Hi!
> >
> > Pretty common task. SCSI initator driver (SCSI miniport or
> SCSI port)
> > with TDI lower edge. iSCSI protocol in separate kernel mode
> DLL. 7-8
> > MB/sec (RAW disk read/write) for 100 Mbit cross-over hubless
> > connection. Is this good or bad? I mean speed. Can anybody tell me
> > what performance boost we’re expected to get when we’ll
> migrate from
> > SCSI miniport that use “tandem SRB completion” to fully deserialzed
> > monolithic SCSI port with TDI called at DISPATCH level?
> Does the time
> > invested into development will cover the resulting
> performace? [Will
> > test with 400 Mbit FirWire and GE soon]
> >
> > Any ideas, iSCSI gurus?
> >
> > Thanks!
> >
> > Anton
> >
> > —
> > You are currently subscribed to ntdev as: xxxxx@alacritech.com To
> > unsubscribe send a blank email to %%email.unsub%%
> >
>
>
>
> —
> You are currently subscribed to ntdev as:
> xxxxx@setengineering.com To unsubscribe send a blank
> email to %%email.unsub%%
>
Yes, we run at 200MB/s with a single initiator against a single
target routinely. It requires a good target and a TCP offload adapter,
however.
----- Original Message -----
From: “Gera Kazakov”
To: “NT Developers Interest List”
Sent: Thursday, July 25, 2002 9:51 AM
Subject: [ntdev] Re: Performance of iSCSI
> > That’s pretty bad. It should be pretty easy to saturate a
> > 100Mb full-duplex link which would give you ~20MB/sec.
>
> Have you ever seen a real initiator/target pair working on such troughput?
> You’d need an almost-zero-latency target and good mix of reads and writes
in
> your test pattern to get this number.
> 7-8, maybe 10Mb/s sounds more realistic.
>
> >
> > What’s your CPU at? It can’t be very high at that kind of
> > throughput unless something is terribly wrong. My guess is
> > this is a latency issue. How many i/o’s do you have outstanding?
> >
> > ----- Original Message -----
> > From: “Anton Kolomyeytsev”
> > To: “NT Developers Interest List”
> > Sent: Thursday, July 25, 2002 9:22 AM
> > Subject: [ntdev] Performance of iSCSI
> >
> >
> > > Hi!
> > >
> > > Pretty common task. SCSI initator driver (SCSI miniport or
> > SCSI port)
> > > with TDI lower edge. iSCSI protocol in separate kernel mode
> > DLL. 7-8
> > > MB/sec (RAW disk read/write) for 100 Mbit cross-over hubless
> > > connection. Is this good or bad? I mean speed. Can anybody tell me
> > > what performance boost we’re expected to get when we’ll
> > migrate from
> > > SCSI miniport that use “tandem SRB completion” to fully deserialzed
> > > monolithic SCSI port with TDI called at DISPATCH level?
> > Does the time
> > > invested into development will cover the resulting
> > performace? [Will
> > > test with 400 Mbit FirWire and GE soon]
> > >
> > > Any ideas, iSCSI gurus?
> > >
> > > Thanks!
> > >
> > > Anton
> > >
> > > —
> > > You are currently subscribed to ntdev as: xxxxx@alacritech.com To
> > > unsubscribe send a blank email to %%email.unsub%%
> > >
> >
> >
> >
> > —
> > You are currently subscribed to ntdev as:
> > xxxxx@setengineering.com To unsubscribe send a blank
> > email to %%email.unsub%%
> >
>
>
> —
> You are currently subscribed to ntdev as: xxxxx@alacritech.com
> To unsubscribe send a blank email to %%email.unsub%%
>
Gera,
I beat 20 Maps in IoMeter, but it was not iSCSI. On a 4 disk JBOD using our
fibre channel card I have hit 80+ MBps using an early version of our virtual
SCSI miniport. Of course that is not iSCSI, and I was able to use megabyte
block sizes and concurrent IO. In a peer to peer, non-SCSI, connection, I
have hit 100+ MBps with a 1Meg buffer and 8 concurrent transfers between
Initiator/Target.
SCSIPORT has real heartburn in a virtual world, and I believe most iSCSI
driver/miniports are going to be virtual. If SCSIPORT controls the hardware
then most likely latency will be minimal, but if the hardware is owned by
other than SCSIPORT (or STORPORT) then latencys will sky rocket as each
transfer to and from the SCSI miniport is going to have a built in quantum
delay. There is no mechanism in either SCSIPORT or STORPORT to queue the DPC
associated with the FDO, and complete an SRB. The SRB is only completed in
that DPC queue. A virtual SCSI miniport, with hardware controlled by an
ethernet driver, can only use a RequestTimer in ScsiPortNotification, and
that incurs a quantum delay. My early SCSI miniport was employing a backdoor
I found to avoid that quantum delay.
It’s to bad the SCSIPORT/STORPORT developers have not found this problem
“remotely interesting”.
–
Gary G. Little
xxxxx@broadstor.com
xxxxx@inland.net
“Gera Kazakov” wrote in message
news:xxxxx@ntdev…
>
> > That’s pretty bad. It should be pretty easy to saturate a
> > 100Mb full-duplex link which would give you ~20MB/sec.
>
> Have you ever seen a real initiator/target pair working on such troughput?
> You’d need an almost-zero-latency target and good mix of reads and writes
in
> your test pattern to get this number.
> 7-8, maybe 10Mb/s sounds more realistic.
>
> >
> > What’s your CPU at? It can’t be very high at that kind of
> > throughput unless something is terribly wrong. My guess is
> > this is a latency issue. How many i/o’s do you have outstanding?
> >
> > ----- Original Message -----
> > From: “Anton Kolomyeytsev”
> > To: “NT Developers Interest List”
> > Sent: Thursday, July 25, 2002 9:22 AM
> > Subject: [ntdev] Performance of iSCSI
> >
> >
> > > Hi!
> > >
> > > Pretty common task. SCSI initator driver (SCSI miniport or
> > SCSI port)
> > > with TDI lower edge. iSCSI protocol in separate kernel mode
> > DLL. 7-8
> > > MB/sec (RAW disk read/write) for 100 Mbit cross-over hubless
> > > connection. Is this good or bad? I mean speed. Can anybody tell me
> > > what performance boost we’re expected to get when we’ll
> > migrate from
> > > SCSI miniport that use “tandem SRB completion” to fully deserialzed
> > > monolithic SCSI port with TDI called at DISPATCH level?
> > Does the time
> > > invested into development will cover the resulting
> > performace? [Will
> > > test with 400 Mbit FirWire and GE soon]
> > >
> > > Any ideas, iSCSI gurus?
> > >
> > > Thanks!
> > >
> > > Anton
> > >
> > > —
> > > You are currently subscribed to ntdev as: xxxxx@alacritech.com To
> > > unsubscribe send a blank email to %%email.unsub%%
> > >
> >
> >
> >
> > —
> > You are currently subscribed to ntdev as:
> > xxxxx@setengineering.com To unsubscribe send a blank
> > email to %%email.unsub%%
> >
>
>
>
Yes, but I mean real-life situation - two regular 100mb netwok cards
connected through at least one switch, regular non-zero-copy TCP stacks on
both sides etc. And a real disk i/o stream which is not symmetrical in
general case. It’s hard to believe that you can reach full wire speed in
such conditions. A half of that is my best bet.
-----Original Message-----
From: Peter Craft [mailto:xxxxx@alacritech.com]
Sent: Thursday, July 25, 2002 11:13 AM
To: NT Developers Interest List
Subject: [ntdev] Re: Performance of iSCSI
Yes, we run at 200MB/s with a single initiator against a
single target routinely. It requires a good target and a TCP
offload adapter, however.
----- Original Message -----
From: “Gera Kazakov”
> To: “NT Developers Interest List”
> Sent: Thursday, July 25, 2002 9:51 AM
> Subject: [ntdev] Re: Performance of iSCSI
>
>
> > > That’s pretty bad. It should be pretty easy to saturate a 100Mb
> > > full-duplex link which would give you ~20MB/sec.
> >
> > Have you ever seen a real initiator/target pair working on such
> > troughput? You’d need an almost-zero-latency target and good mix of
> > reads and writes
> in
> > your test pattern to get this number.
> > 7-8, maybe 10Mb/s sounds more realistic.
> >
> > >
> > > What’s your CPU at? It can’t be very high at that kind of
> > > throughput unless something is terribly wrong. My guess
> is this is
> > > a latency issue. How many i/o’s do you have outstanding?
> > >
> > > ----- Original Message -----
> > > From: “Anton Kolomyeytsev”
> > > To: “NT Developers Interest List”
> > > Sent: Thursday, July 25, 2002 9:22 AM
> > > Subject: [ntdev] Performance of iSCSI
> > >
> > >
> > > > Hi!
> > > >
> > > > Pretty common task. SCSI initator driver (SCSI miniport or
> > > SCSI port)
> > > > with TDI lower edge. iSCSI protocol in separate kernel mode
> > > DLL. 7-8
> > > > MB/sec (RAW disk read/write) for 100 Mbit cross-over hubless
> > > > connection. Is this good or bad? I mean speed. Can
> anybody tell me
> > > > what performance boost we’re expected to get when we’ll
> > > migrate from
> > > > SCSI miniport that use “tandem SRB completion” to fully
> > > > deserialzed monolithic SCSI port with TDI called at DISPATCH
> > > > level?
> > > Does the time
> > > > invested into development will cover the resulting
> > > performace? [Will
> > > > test with 400 Mbit FirWire and GE soon]
> > > >
> > > > Any ideas, iSCSI gurus?
> > > >
> > > > Thanks!
> > > >
> > > > Anton
> > > >
> > > > —
> > > > You are currently subscribed to ntdev as:
> xxxxx@alacritech.com To
> > > > unsubscribe send a blank email to %%email.unsub%%
> > > >
> > >
> > >
> > >
> > > —
> > > You are currently subscribed to ntdev as:
> > > xxxxx@setengineering.com To unsubscribe send a
> blank email to
> > > %%email.unsub%%
> > >
> >
> >
> > —
> > You are currently subscribed to ntdev as: xxxxx@alacritech.com To
> > unsubscribe send a blank email to %%email.unsub%%
> >
>
>
>
> —
> You are currently subscribed to ntdev as:
> xxxxx@setengineering.com To unsubscribe send a blank
> email to %%email.unsub%%
>